Unlike closed-vocabulary 3D instance segmentation that is often trained end-to-end, open-vocabulary 3D instance segmentation (OV-3DIS) often leverages vision-language models (VLMs) to generate 3D instance proposals and classify them. While various concepts have been proposed from existing research, we observe that these individual concepts are not mutually exclusive but complementary. In this paper, we propose a new state-of-the-art solution for OV-3DIS by carefully designing a recipe to combine the concepts together and refining them to address key challenges. Our solution follows the two-stage scheme: 3D proposal generation and instance classification. We employ robust 3D tracking-based proposal aggregation to generate 3D proposals and remove overlapped or partial proposals by iterative merging/removal. For the classification stage, we replace the standard CLIP model with Alpha-CLIP, which incorporates object masks as an alpha channel to reduce background noise and obtain object-centric representation. Additionally, we introduce the standardized maximum similarity (SMS) score to normalize text-to-proposal similarity, effectively filtering out false positives and boosting precision. Our framework achieves state-of-the-art performance on ScanNet200 and S3DIS across all AP and AR metrics, even surpassing an end-to-end closed-vocabulary method.

We investigate the problem of teaching a robot manipulator to perform dynamic non-prehensile object transport, also known as the ‘robot waiter’ task, from a limited set of real-world demonstrations. We propose an approach that combines batch reinforcement learning (RL) with modelpredictive control (MPC) by pretraining an ensemble of value functions from demonstration data, and utilizing them online within an uncertainty-aware MPC scheme to ensure robustness to limited data coverage. Our approach is straightforward to integrate with off-the-shelf MPC frameworks and enables learning solely from task space demonstrations with sparsely labeled transitions, while leveraging MPC to ensure smooth joint space motions and constraint satisfaction. We validate the proposed approach through extensive simulated and real-world experiments on a Franka Panda robot performing the robot waiter task and demonstrate robust deployment of value functions learned from 50 – 100 demonstrations. Furthermore, our approach enables generalization to novel objects not seen during training and can improve upon suboptimal demonstrations. We believe that such a framework can reduce the burden of providing extensive demonstrations and facilitate rapid training of robot manipulators to perform non-prehensile manipulation tasks. Project videos and supplementary material can be found at: https://sites.google.com/view/cvmpc

We focus on agile, continuous, and terrain-adaptive jumping of quadrupedal robots in discontinuous terrains such as stairs and stepping stones. Unlike single-step jumping, continuous jumping requires accurately executing highly dynamic motions over long horizons, which is challenging for existing approaches. To accomplish this task, we design a hierarchical learning and control framework, which consists of a learned heightmap predictor for robust terrain perception, a reinforcement-learning-based centroidal-level motion policy for versatile and terrain-adaptive planning, and a low-level model-based leg controller for accurate motion tracking. In addition, we minimize the sim-to-real gap by accurately modeling the hardware characteristics. Our framework enables a Unitree Go1 robot to perform agile and continuous jumps on human-sized stairs and sparse stepping stones, for the first time to the best of our knowledge. In particular, the robot can cross two stair steps in each jump and completes a 3.5m long, 2.8m high, 14-step staircase in 4.5 seconds. Moreover, the same policy outperforms baselines in various other parkour tasks, such as jumping over single horizontal or vertical discontinuities. Experiment videos can be found at https://yxyang.github.io/jumping_cod/.

Training reinforcement learning (RL) agents directly from high-dimensional image observations continues to be a challenging problem. Recent line of work on behavioral distances proposes to learn representations that encode behavioral similarities quantified by the bisimulation metric. By learning an isometric mapping to a lower dimensional space that preserves this metric, such methods attempt to learn representations that group together functionally similar states. However, such an isometric mapping may not exist, making the learning objective ill-defined. We propose an alternative objective that allows distortions in long-range distances, while preserving local metric structure – inducing representations that highlight natural clusters in the state space. This leads to new representations, which we term Behavioral Eigenmaps (BeigeMaps), corresponding to the eigenfunctions of similarity kernels induced by behavioral distances. We empirically demonstrate that when added as a drop-in modification, BeigeMaps improve the policy performance of prior behavioral distance based RL algorithms.

Reliable estimation of terrain traversability is critical for the successful deployment of autonomous systems in wild, outdoor environments. Given the lack of large-scale annotated datasets for off-road navigation, strictly-supervised learning approaches remain limited in their generalization ability. To this end, we introduce a novel, image-based self-supervised learning method for traversability prediction, leveraging a state-of-the-art vision foundation model for improved out-of-distribution performance. Our method employs contrastive representation learning using both human driving data and instance-based segmentation masks during training. We show that this simple, yet effective, technique drastically outperforms recent methods in predicting traversability for both on- and off-trail driving scenarios. We compare our method with recent baselines on both a common benchmark as well as our own datasets, covering a diverse range of outdoor environments and varied terrain types. We also demonstrate the compatibility of resulting costmap predictions with a model-predictive controller. Finally, we evaluate our approach on zero- and few-shot tasks, demonstrating unprecedented performance for generalization to new environments. Videos and additional material can be found here: https://sites.google.com/view/visual-traversability-learning.

A major challenge in robotics is to design robust policies which enable complex and agile behaviors in the real world. On one end of the spectrum, we have model-free reinforcement learning (MFRL), which is incredibly flexible and general but often results in brittle policies. In contrast, model predictive control (MPC) continually re-plans at each time step to remain robust to perturbations and model inaccuracies. However, despite its real-world successes, MPC often under-performs the optimal strategy. This is due to model quality, myopic behavior from short planning horizons, and approximations due to computational constraints. And even with a perfect model and enough compute, MPC can get stuck in bad local optima, depending heavily on the quality of the optimization algorithm. To this end, we propose Deep Model Predictive Optimization (DMPO), which learns the inner-loop of an MPC optimization algorithm directly via experience, specifically tailored to the needs of the control problem. We evaluate DMPO on a real quadrotor agile trajectory tracking task, on which it improves performance over a baseline MPC algorithm for a given computational budget. It can outperform the best MPC algorithm by up to 27% with fewer samples and an end-to-end policy trained with MFRL by 19%. Moreover, because DMPO requires fewer samples, it can also achieve these benefits with 4.3× less memory. When we subject the quadrotor to turbulent wind fields with an attached drag plate, DMPO can adapt zero-shot while still outperforming all baselines. Additional results can be found at https://tinyurl.com/mr2ywmnw.

Self-recovery is a critical capability for robust, agile robots operating in the real world. Given truly challenging terrain, it is nearly inevitable that, at some point, the robot will fail and subsequently need to recover if it is to continue its task. One critical subset of recovery is standing back up after falling down (aka "self-righting"), an essential early milestone for babies learning to walk, and an existential capability for animals. While some robots can be designed with multiple orientations for mobility, most seeking to affect the world would significantly benefit from planners/policies that facilitate self-righting whenever possible.In this work, we present a series of challenges that outline why recovery in the wild is difficult. We then present a set of benchmark policies trained in simulation using deep reinforcement learning (RL) and the Student-Teacher approach. Finally, we evaluate the performance of these policies on a set of benchmark contexts in simulation, and provide baseline validation on a physical robot.

We focus on the problem of long-range dynamic replanning for off-road autonomous vehicles, where a robot plans paths through a previously unobserved environment while continuously receiving noisy local observations. An effective approach for planning under sensing uncertainty is determinization, where one converts a stochastic world into a deterministic one and plans under this simplification. This makes the planning problem tractable, but the cost of following the planned path in the real world may be different than in the determinized world. This causes collisions if the determinized world optimistically ignores obstacles, or causes unnecessarily long routes if the determinized world pessimistically imagines more obstacles. We aim to be robust to uncertainty over potential worlds while still achieving the efficiency benefits of determinization. We evaluate algorithms for dynamic replanning on a large real-world dataset of challenging long-range planning problems from the DARPA RACER program. Our method, Dynamic Replanning via Evaluating and Aggregating Multiple Samples (DREAMS), outperforms other determinization-based approaches in terms of combined traversal time and collision cost. https://sites.google.com/cs.washington.edu/dreams/

Quadrupedal robots have emerged as versatile agents capable of locomoting and manipulating in complex environments. Traditional designs typically rely on the robot’s inherent body parts or incorporate top-mounted arms for manipulation tasks. However, these configurations may limit the robot’s operational dexterity, efficiency, and adaptability, particularly in cluttered or constrained spaces. In this work, we present LocoMan, a dexterous quadrupedal robot with a novel morphology to perform versatile manipulation in diverse constrained environments. By equipping a Unitree Go1 robot with two low-cost and lightweight modular 3-DoF loco-manipulators on its front calves, LocoMan leverages the combined mobility and functionality of the legs and grippers for complex manipulation tasks that require precise 6D positioning of the end effector in a wide workspace. To harness the loco-manipulation capabilities of LocoMan, we introduce a unified control framework that extends the whole-body controller (WBC) to integrate the dynamics of loco-manipulators. Through experiments, we validate that the proposed whole-body controller can accurately and stably follow desired 6D trajectories of the end effector and torso, which, when combined with the large workspace from our design, facilitates a diverse set of challenging dexterous loco-manipulation tasks in confined spaces, such as opening doors, plugging into sockets, picking objects in narrow and low-lying spaces, and bimanual manipulation.

Despite recent advancements in deep learning, deep neural networks continue to suffer from performance degradation when applied to new data that differs from training data. Test-time adaptation (TTA) aims to address this challenge by adapting a model to unlabeled data at test time. TTA can be applied to pretrained networks without modifying their training procedures, enabling them to utilize a well-formed source distribution for adaptation. One possible approach is to align the representation space of test samples to the source distribution (i.e., feature alignment). However, performing feature alignment in TTA is especially challenging in that access to labeled source data is restricted during adaptation. That is, a model does not have a chance to learn test data in a class-discriminative manner, which was feasible in other adaptation tasks (e.g., unsupervised domain adaptation) via supervised losses on the source data. Based on this observation, we propose a simple yet effective feature alignment loss, termed as Class-Aware Feature Alignment (CAFA), which simultaneously 1) encourages a model to learn target representations in a class-discriminative manner and 2) effectively mitigates the distribution shifts at test time. Our method does not require any hyper-parameters or additional losses, which are required in previous approaches. We conduct extensive experiments on 6 different datasets and show our proposed method consistently outperforms existing baselines.

We introduce LiDAR-UDA, a novel two-stage self-training-based Unsupervised Domain Adaptation (UDA) method for LiDAR segmentation. Existing self-training methods use a model trained on labeled source data to generate pseudo labels for target data and refine the predictions via fine-tuning the network on the pseudo labels. These methods suffer from domain shifts caused by different LiDAR sensor configurations in the source and target domains. We propose two techniques to reduce sensor discrepancy and improve pseudo label quality: 1) Li-DAR beam subsampling, which simulates different LiDAR scanning patterns by randomly dropping beams; 2) cross-frame ensembling, which exploits temporal consistency of consecutive frames to generate more reliable pseudo labels. Our method is simple, generalizable, and does not incur any extra inference cost. We evaluate our method on several public LiDAR datasets and show that it outperforms the state-of-the-art methods by more than 3.9% mIoU on average for all scenarios. Code will be available at https://github.com/JHLee0513/lidar_uda.

We study a new paradigm for sequential decision making, called offline policy learning from observations (PLfO). Offline PLfO aims to learn policies using datasets with substandard qualities: 1) only a subset of trajectories is labeled with rewards, 2) labeled trajectories may not contain actions, 3) labeled trajectories may not be of high quality, and 4) the data may not have full coverage. Such imperfection is common in real-world learning scenarios, and offline PLfO encompasses many existing offline learning setups, including offline imitation learning (IL), offline IL from observations (ILfO), and offline reinforcement learning (RL). In this work, we present a generic approach to offline PLfO, called Modality-agnostic Adversarial Hypothesis Adaptation for Learning from Observations (MAHALO). Built upon the pessimism concept in offline RL, MAHALO optimizes the policy using a performance lower bound that accounts for uncertainty due to the dataset’s insufficient coverage. We implement this idea by adversarially training data-consistent critic and reward functions, which forces the learned policy to be robust to data deficiency. We show that MAHALO consistently outperforms or matches specialized algorithms across a variety of offline PLfO tasks in theory and experiments. Our code is available at https://github.com/AnqiLi/mahalo.

We present Neural Contact Fields, a method that brings together neural fields and tactile sensing to address the problem of tracking extrinsic contact between object and environment. Knowing where the external contact occurs is a first step towards methods that can actively control it in facilitating downstream manipulation tasks. Prior work for localizing environmental contacts typically assume a contact type (e.g. point or line), does not capture contact/no-contact transitions, and only works with basic geometric-shaped objects. Neural Contact Fields are the first method that can track arbitrary multi-modal extrinsic contacts without making any assumptions about the contact type. Our key insight is to estimate the probability of contact for any 3D point in the latent space of object's shapes, given vision-based tactile inputs that sense the local motion resulting from the external contact. In experiments, we find that Neural Contact Fields are able to localize multiple contact patches without making any assumptions about the geometry of the contact, and capture contact/no-contact transitions for known categories of objects with unseen shapes in unseen environment configurations. In addition to Neural Contact Fields, we also release our YCB-Extrinsic-Contact dataset of simulated extrinsic contact interactions to enable further research in this area. Project page: https://github.com/carolinahiguera/NCF

Autocurricular training is an important sub-area of multi-agent reinforcement learning (MARL) that allows multiple agents to learn emergent skills in an unsupervised co-evolving scheme. The robotics community has experimented auto-curricular training with physically grounded problems, such as robust control and interactive manipulation tasks. However, the asymmetric nature of these tasks makes the generation of sophisticated policies challenging. Indeed, the asymmetry in the environment may implicitly or explicitly provide an advantage to a subset of agents which could, in turn, lead to a low-quality equilibrium. This paper proposes a novel game-theoretic algorithm, Stackelberg Multi-Agent Deep Deterministic Policy Gradient (ST-MADDPG), which formulates a two-player MARL problem as a Stackelberg game with one player as the ‘leader’ and the other as the ‘follower’ in a hierarchical interaction structure wherein the leader has an advantage. We first demonstrate that the leader's advantage from ST-MADDPG can be used to alleviate the inherent asymmetry in the environment. By exploiting the leader's advantage, ST-MADDPG improves the quality of a co-evolution process and results in more sophisticated and complex strategies that work well even against an unseen strong opponent.

We propose a novel model-based offline Reinforcement Learning (RL) framework, called Adversarial Model for Offline Reinforcement Learning (ARMOR), which can robustly learn policies to improve upon an arbitrary reference policy regardless of data coverage. ARMOR is designed to optimize policies for the worst-case performance relative to the reference policy through adversarially training a Markov decision process model. In theory, we prove that ARMOR, with a well-tuned hyperparameter, can compete with the best policy within data coverage when the reference policy is supported by the data. At the same time, ARMOR is robust to hyperparameter choices: the policy learned by ARMOR, with any admissible hyperparameter, would never degrade the performance of the reference policy, even when the reference policy is not covered by the dataset. To validate these properties in practice, we design a scalable implementation of ARMOR, which by adversarial training, can optimize policies without using model ensembles in contrast to typical model-based methods. We show that ARMOR achieves competent performance with both state-of-the-art offline model-free and model-based RL algorithms and can robustly improve the reference policy over various hyperparameter choices.

Kernel herding is a deterministic sampling algorithm designed to draw ‘super samples' from probability distributions when provided with their kernel mean embeddings in a reproducing kernel Hilbert space (RKHS). Empirical expectations of functions in the RKHS formed using these super samples tend to converge even faster than random sampling from the true distribution itself. Standard implementations of kernel herding have been restricted to sampling over flat Euclidean spaces, which is not ideal for applications such as robotics where more general Riemannian manifolds may be appropriate. We propose to adapt kernel herding to Riemannian manifolds by (1) using geometry-aware kernels that incorporate the appropriate distance metric for the manifold and (2) using Riemannian optimization to constrain herded samples to lie on the manifold. We evaluate our approach on problems involving various manifolds commonly used in robotics including the SO(3) manifold of rotation matrices, the spherical manifold used to encode unit quaternions, and the manifold of symmetric positive definite matrices. We demonstrate that our approach outperforms existing alternatives on the task of resampling from empirical distributions of weighted particles, a problem encountered in applications such as particle filtering. We also demonstrate how Riemannian kernel herding can be used as part of the kernel recursive approximate Bayesian computation algorithm to estimate parameters of black-box simulators, including inertia matrices of an Adroit robot hand simulator. Our results confirm that exploiting geometric information through our approach to kernel herding yields better results than alternatives including standard kernel herding with heuristic projections.

Efficient and reliable generation of global path plans are necessary for safe execution and deployment of autonomous systems. In order to generate planning graphs which adequately resolve the topology of a given environment, many sampling-based motion planners resort to coarse, heuristically-driven strategies which often fail to generalize to new and varied surroundings. Further, many of these approaches are not designed to contend with partial-observability. We posit that such uncertainty in environment geometry can, in fact, help drive the sampling process in generating feasible, and probabilistically-safe planning graphs. We propose a method for Probabilistic Roadmaps which relies on particle-based Variational Inference to efficiently cover the posterior distribution over feasible regions in configuration space. Our approach, Stein Variational Probabilistic Roadmap (SV-PRM), results in sample-efficient generation of planning-graphs and large improvements over traditional sampling approaches. We demonstrate the approach on a variety of challenging planning problems, including real-world probabilistic occupancy maps and high-dof manipulation problems common in robotics. Video, additional material and results can be found here: https://sites.google.com/view/stein-prm.

Sampling-based Model Predictive Control (MPC) is a flexible control framework that can reason about non-smooth dynamics and cost functions. Recently, significant work has focused on the use of machine learning to improve the performance of MPC, often through learning or fine-tuning the dynamics or cost function. In contrast, we focus on learning to optimize more effectively. In other words, to improve the update rule within MPC. We show that this can be particularly useful in sampling-based MPC, where we often wish to minimize the number of samples for computational reasons. Unfortunately, the cost of computational efficiency is a reduction in performance; fewer samples results in noisier updates. We show that we can contend with this noise by learning how to update the control distribution more effectively and make better use of the few samples that we have. Our learned controllers are trained via imitation learning to mimic an expert which has access to substantially more samples. We test the efficacy of our approach on multiple simulated robotics tasks in sample-constrained regimes and demonstrate that our approach can outperform a MPC controller with the same number of samples.

Motion optimization is an effective framework for generating smooth and safe trajectories for robotic manipulation tasks. However, it suffers from local optima that hinder its applicability, especially for multi-objective tasks. In this paper, we study this problem in light of the integration of Energy-Based Models (EBM) as guiding priors in motion optimization. EBMs are probabilistic models with unnormalized energy functions that represent expressive multimodal distributions. Due to their implicit nature, EBMs can easily be integrated as data-driven factors or initial sampling distributions in the motion optimization problem. This work presents a set of necessary modeling and algorithmic choices to effectively learn and integrate EBMs into motion optimization. We present a set of EBM architectures for learning generalizable distributions over trajectories that are important for the subsequent deployment of EBMs. Moreover, we investigate the benefit of including smoothness regularization in the learning process to improve motion optimization. In addition to gradient-based solvers, we also propose a stochastic method for trajectory optimization with learned EBMs. We provide extensive empirical results in a set of representative tasks against competitive baselines that demonstrate the superiority of EBMs as priors in motion optimization scaling up to 7 -dof robot pouring that can be easily transferred to the real robotic system. Videos and additional details are available at https://sites.google.com/view/implicit-priors.

Detecting novel objects from few examples has become an emerging topic in computer vision recently. However, current methods need fully annotated training images to learn new object categories which limits their applicability in real world scenarios such as field robotics. In this work, we propose a probabilistic multiple-instance learning approach for few-shot Common Object Localization (COL) and few-shot Weakly Supervised Object Detection (WSOD). In these tasks, only image-level labels, which are much cheaper to acquire, are available. We find that operating on features extracted from the last layer of a pretrained Faster-RCNN is more effective compared to previous episodic learning based few-shot COL methods. Our model simultaneously learns the distribution of the novel objects and localizes them via expectation-maximization steps. As a probabilistic model, we employ von Mises-Fisher (vMF) distribution which captures the semantic information better than Gaussian distribution when applied to the pre-trained embedding space. When the novel objects are localized, we utilize them to learn a linear appearance model to detect novel classes in new images. Our extensive experiments show that the proposed method, despite being simple, outperforms strong baselines in few-shot COL and WSOD, as well as large-scale WSOD tasks.

Modeling joint probability distributions over sequences has been studied from many perspectives. The physics community developed matrix product states, a tensor-train decomposition for probabilistic modeling, motivated by the need to tractably model many-body systems. But similar models have also been studied in the stochastic processes and weighted automata literature, with little work on how these bodies of work relate to each other. We address this gap by showing how stationary or uniform versions of popular quantum tensor network models have equivalent representations in the stochastic processes and weighted automata literature, in the limit of infinitely long sequences. We demonstrate several equivalence results between models used in these three communities: (i) uniform variants of matrix product states, Born machines and locally purified states from the quantum tensor networks literature, (ii) predictive state representations, hidden Markov models, norm-observable operator models and hidden quantum Markov models from the stochastic process literature, and (iii) stochastic weighted automata, probabilistic automata and quadratic automata from the formal languages literature. Such connections may open the door for results and methods developed in one area to be applied in another.

Many sequential decision problems involve finding a policy that maximizes total reward while obeying safety constraints. Although much recent research has focused on the development of safe reinforcement learning (RL) algorithms that produce a safe policy after training, ensuring safety during training as well remains an open problem. A fundamental challenge is performing exploration while still satisfying constraints in an unknown Markov decision process (MDP). In this work, we address this problem for the chance-constrained setting.We propose a new algorithm, SAILR, that uses an intervention mechanism based on advantage functions to keep the agent safe throughout training and optimizes the agent’s policy using off-the-shelf RL algorithms designed for unconstrained MDPs. Our method comes with strong guarantees on safety during "both" training and deployment (i.e., after training and without the intervention mechanism) and policy performance compared to the optimal safety-constrained policy. In our experiments, we show that SAILR violates constraints far less during training than standard safe RL and constrained MDP approaches and converges to a well-performing policy that can be deployed safely without intervention. Our code is available at https://github.com/nolanwagener/safe_rl.

Model Predictive Control (MPC) is a classic tool for optimal control of complex, real-world systems. Although it has been successfully applied to a wide range of challenging tasks in robotics, it is fundamentally limited by the prediction horizon, which, if too short, will result in myopic decisions. Recently, several papers have suggested using a learned value function as the terminal cost for MPC. If the value function is accurate, it effectively allows MPC to reason over an infinite horizon. Unfortunately, Reinforcement Learning (RL) solutions to value function approximation can be difficult to realize for robotics tasks. In this paper, we suggest a more efficient method for value function approximation that applies to goal-directed problems, like reaching and navigation. In these problems, MPC is often formulated to track a path or trajectory returned by a planner. However, this strategy is brittle in that unexpected perturbations to the robot will require replanning, which can be costly at runtime. Instead, we show how the intermediate data structures used by modern planners can be interpreted as an approximate value function. We show that that this value function can be used by MPC directly, resulting in more efficient and resilient behavior at runtime.

Billions of people use chopsticks, a simple yet versatile tool, for fine manipulation of everyday objects. The small, curved, and slippery tips of chopsticks pose a challenge for picking up small objects, making them a suitably complex test case. This paper leverages human demonstrations to develop an autonomous chopsticks-equipped robotic manipulator. Due to the lack of accurate models for fine manipulation, we explore model-free imitation learning, which traditionally suffers from the covariate shift phenomenon that causes poor generalization. We propose two approaches to reduce covariate shift, neither of which requires access to an interactive expert or a model, unlike previous approaches. First, we alleviate singlestep prediction errors by applying an invariant operator to increase the data support at critical steps for grasping. Second, we generate synthetic corrective labels by adding bounded noise and combining parametric and non-parametric methods to prevent error accumulation. We demonstrate our methods on a real chopstick-equipped robot that we built, and observe the agent’s success rate increase from 37.3% to 80%, which is comparable to the human expert performance of 82.6%.

Generating robot motion that fulfills multiple tasks simultaneously is challenging due to the geometric constraints imposed on the robot. In this paper, we propose to solve multi-task problems through learning structured policies from human demonstrations. Our structured policy is inspired by RMPflow, a framework for combining subtask policies on different spaces. The policy structure provides the user an interface to 1) specifying the spaces that are directly relevant to the completion of the tasks, and 2) designing policies for certain tasks that do not need to be learned. We derive an end-to-end learning objective that is suitable for the multi-task problem, emphasizing the distance between generated motions and demonstrations measured on task spaces. Furthermore, the motion generated from the learned policy class is guaranteed to be stable. We validate the effectiveness of our proposed learning framework through qualitative and quantitative evaluations on three robotic tasks on a 7-DOF Rethink Sawyer robot.

Online imitation learning (IL) is an algorithmic framework that leverages interactions with expert policies for efficient policy optimization. Here policies are optimized by performing online learning on a sequence of loss functions that encourage the learner to mimic expert actions, and if the online learning has no regret, the agent can provably learn an expert-like policy. Online IL has demonstrated empirical successes in many applications and interestingly, its policy improvement speed observed in practice is usually much faster than existing theory suggests. In this work, we provide an explanation of this phenomenon. Let $\xi$ denote the policy class bias and assume the online IL loss functions are convex, smooth, and non-negative. We prove that, after $N$ rounds of online IL with stochastic feedback, the policy improves in $\tilde{O}(1/N + \sqrt{\xi/N})$ in both expectation and high probability. In other words, we show that adopting a sufficiently expressive policy class in online IL has two benefits: both the policy improvement speed increases and the performance bias decreases.

Generating robot motion for multiple tasks in dynamic environments is challenging, requiring an algorithm to respond reactively while accounting for complex nonlinear relationships between tasks. In this article, we develop a novel policy synthesis algorithm, Riemannian motion policy (RMP)flow, based on geometrically consistent transformations of RMPs. RMPs are a class of reactive motion policies that parameterize non-Euclidean behaviors as dynamical systems in intrinsically nonlinear task spaces. Given a set of RMPs designed for individual tasks, RMPflow can combine these policies to generate an expressive global policy, while simultaneously exploiting sparse structure for computational efficiency. We study the geometric properties of RMPflow and provide sufficient conditions for stability. Finally, we experimentally demonstrate that accounting for the natural Riemannian geometry of task policies can simplify classically difficult problems, such as planning through the clutter on high-degree-of-freedom manipulation systems.Note to Practitioners—Requirements on safety and responsiveness for collaborative robots have driven a need for new ideas in control design that bridge between standard objectives in low-level control (such as trajectory tracking) and high-level behavioral objectives (such as collision avoidance) often relegated to planning systems. Modern results from geometric control, which promise stable controllers that can smoothly and safely transition between many behavioral tasks, therefore, become highly relevant. However, for years, this field has remained inaccessible due to its mathematical complexity. This article aims to: 1) make those ideas accessible to robotics and control experts by recasting them in a concrete algorithmic framework amenable to controller design and 2) to additionally generalize them to better satisfy the specific needs of robotic behavior generation. Our experiments demonstrate that the resulting controllers can engender natural behavior that adapts instantaneously to changing surroundings with zero planning while performing manipulation tasks. The framework is gaining traction within the robotics community, finding increasing application in areas, such as autonomous navigation, tactile servoing, and multi-agent systems. Future research will address learning these controllers from data to simplify that process of design and tuning, which at present can require experience.

Extending classical probabilistic reasoning using the quantum mechanical view of probability has been of recent interest, particularly in the development of hidden quantum Markov models (HQMMs) to model stochastic processes. However, there has been little progress in characterizing the expressiveness of such models and learning them from data. We tackle these problems by showing that HQMMs are a special subclass of the general class of observable operator models (OOMs) that do not suffer from the negative probability problem by design. We also provide a feasible retraction-based learning algorithm for HQMMs using constrained gradient descent on the Stiefel manifold of model parameters. We demonstrate that this approach is faster and scales to larger models than previous learning algorithms.

We present a reduction from reinforcement learning (RL) to no-regret online learning based on the saddle-point formulation of RL, by which "any" online algorithm with sublinear regret can generate policies with provable performance guarantees. This new perspective decouples the RL problem into two parts: regret minimization and function approximation. The first part admits a standard online-learning analysis, and the second part can be quantified independently of the learning algorithm. Therefore, the proposed reduction can be used as a tool to systematically design new RL algorithms. We demonstrate this idea by devising a simple RL algorithm based on mirror descent and the generative-model oracle. For any $\gamma$-discounted tabular RL problem, with probability at least $1-\delta$, it learns an $\epsilon$-optimal policy using at most $\tilde{O}\left(\frac{|\SS||Å|\log(\frac{1}{\delta})}{(1-\gamma)^4\epsilon^2}\right)$ samples. Furthermore, this algorithm admits a direct extension to linearly parameterized function approximators for large-scale applications, with computation and sample complexities independent of $|\SS|$,$|Å|$, though at the cost of potential approximation bias.

Online learning is a powerful tool for analyzing iterative algorithms. However, the classic adversarial setup fails to capture regularity that can exist in practice. Motivated by this observation, we establish a new setup, called Continuous Online Learning (COL), where the gradient of online loss function changes continuously across rounds with respect to the learner’s decisions. We show that COL appropriately describes many interesting applications, from general equilibrium problems (EPs) to optimization in episodic MDPs. Using this new setup, we revisit the difficulty of sublinear dynamic regret. We prove a fundamental equivalence between achieving sublinear dynamic regret in COL and solving certain EPs. With this insight, we offer conditions for efficient algorithms that achieve sublinear dynamic regret, even when the losses are chosen adaptively without any a priori variation budget. Furthermore, we show for COL a reduction from dynamic regret to both static regret and convergence in the associated EP, allowing us to analyze the dynamic regret of many existing algorithms.

Modern trajectory optimization based approaches to motion planning are fast, easy to implement, and effective on a wide range of robotics tasks. However, trajectory optimization algorithms have parameters that are typically set in advance (and rarely discussed in detail). Setting these parameters properly can have a significant impact on the practical performance of the algorithm, sometimes making the difference between finding a feasible plan or failing at the task entirely. We propose a method for leveraging past experience to learn how to automatically adapt the parameters of Gaussian Process Motion Planning (GPMP) algorithms. Specifically, we propose a differentiable extension to the GPMP2 algorithm, so that it can be trained end-to-end from data. We perform several experiments that validate our algorithm and illustrate the benefits of our proposed learning-based approach to motion planning.

Learning from offline task demonstrations is a problem of great interest in robotics. For simple short-horizon manipulation tasks with modest variation in task instances, offline learning from a small set of demonstrations can produce controllers that successfully solve the task. However, leveraging a fixed batch of data can be problematic for larger datasets and longer-horizon tasks with greater variations. The data can exhibit substantial diversity and consist of suboptimal solution approaches. In this paper, we propose Implicit Reinforcement without Interaction at Scale (IRIS), a novel framework for learning from large-scale demonstration datasets. IRIS factorizes the control problem into a goal-conditioned low-level controller that imitates short demonstration sequences and a high-level goal selection mechanism that sets goals for the low-level and selectively combines parts of suboptimal solutions leading to more successful task completions. We evaluate IRIS across three datasets, including the RoboTurk Cans dataset collected by humans via crowdsourcing, and show that performant policies can be learned from purely offline learning. Additional results at https://sites.google.com/stanford.edu/iris/.

Effective human-robot collaboration requires informed anticipation. The robot must anticipate the human's actions, but also react quickly and intuitively when its predictions are wrong. The robot must plan its actions to account for the human's own plan, with the knowledge that the human's behavior will change based on what the robot actually does. This cyclical game of predicting a human's future actions and generating a corresponding motion plan is extremely difficult to model using standard techniques. In this work, we describe a novel Model Predictive Control (MPC)-based framework for finding optimal trajectories in a collaborative, multi-agent setting, in which we simultaneously plan for the robot while predicting the actions of its external collaborators. We use human-robot handovers to demonstrate that with a strong model of the collaborator, our framework produces fluid, reactive human-robot interactions in novel, cluttered environments. Our method efficiently generates coordinated trajectories, and achieves a high success rate in handover, even in the presence of significant sensor noise.

Predicting calibrated confidence scores for multi-class deep networks is important for avoiding rare but costly mistakes. A common approach is to learn a post-hoc calibration function that transforms the output of the original network into calibrated confidence scores while maintaining the network's accuracy. However, previous post-hoc calibration techniques work only with simple calibration functions, potentially lacking sufficient representation to calibrate the complex function landscape of deep networks. In this work, we aim to learn general post-hoc calibration functions that can preserve the top-k predictions of any deep network. We call this family of functions intra order-preserving functions. We propose a new neural network architecture that represents a class of intra order-preserving functions by combining common neural network components. Additionally, we introduce order-invariant and diagonal sub-families, which can act as regularization for better generalization when the training data size is small. We show the effectiveness of the proposed method across a wide range of datasets and classifiers. Our method outperforms state-of-the-art post-hoc calibration methods, namely temperature scaling and Dirichlet calibration, in several evaluation metrics for the task.

Sample efficiency is critical in solving real-world reinforcement learning problems where agent-environment interactions can be costly. Imitation learning from expert advice has proved to be an effective strategy for reducing the number of interactions required to train a policy. Online imitation learning, which interleaves policy evaluation and policy optimization, is a particularly effective technique with provable performance guarantees. In this work, we seek to further accelerate the convergence rate of online imitation learning, thereby making it more sample efficient. We propose two model-based algorithms inspired by Follow-the-Leader (FTL) with prediction: MoBIL-VI based on solving variational inequalities and MoBIL-Prox based on stochastic first-order updates. These two methods leverage a model to predict future gradients to speed up policy learning. When the model oracle is learned online, these algorithms can provably accelerate the best known convergence rate up to an order. Our algorithms can be viewed as a generalization of stochastic Mirror-Prox (Juditsky et al., 2011), and admit a simple constructive FTL-style analysis of performance.

Bilevel optimization has been recently revisited for designing and analyzing algorithms in hyperparameter tuning and meta learning tasks. However, due to its nested structure, evaluating exact gradients for high-dimensional problems is computationally challenging. One heuristic to circumvent this difficulty is to use the approximate gradient given by performing truncated back-propagation through the iterative optimization procedure that solves the lower-level problem. Although promising empirical performance has been reported, its theoretical properties are still unclear. In this paper, we analyze the properties of this family of approximate gradients and establish sufficient conditions for convergence. We validate this on several hyperparameter tuning and meta learning tasks. We find that optimization with the approximate gradient computed using few-step back-propagation often performs comparably to optimization with the exact gradient, while requiring far less memory and half the computation time.

Given a collection of bags where each bag is a set of images, our goal is to select one image from each bag such that the selected images are from the same object class. We model the selection as an energy minimization problem with unary and pairwise potential functions. Inspired by recent few-shot learning algorithms, we propose an approach to learn the potential functions directly from the data. Furthermore, we propose a fast greedy inference algorithm for energy minimization. We evaluate our approach on few-shot common object recognition as well as object co-localization tasks. Our experiments show that learning the pairwise and unary terms greatly improves the performance of the model over several well-known methods for these tasks. The proposed greedy optimization algorithm achieves performance comparable to state-of-the-art structured inference algorithms while being ~10 times faster.

We study Imitation Learning (IL) from Observations alone (ILFO) in large-scale MDPs. While most IL algorithms rely on an expert to directly provide actions to the learner, in this setting the expert only supplies sequences of observations. We design a new model-free algorithm for ILFO, Forward Adversarial Imitation Learning (FAIL), which learns a sequence of time-dependent policies by minimizing an Integral Probability Metric between the observation distributions of the expert policy and the learner. FAIL provably learns a near-optimal policy with a number of samples that is polynomial in all relevant parameters but independent of the number of unique observations. The resulting theory extends the domain of provably sample efficient learning algorithms beyond existing results that typically only consider tabular RL settings or settings that require access to a near-optimal reset distribution. We also demonstrate the efficacy ofFAIL on multiple OpenAI Gym control tasks.

We present a predictor-corrector framework, called PicCoLO, that can transform a first-order model-free reinforcement or imitation learning algorithm into a new hybrid method that leverages predictive models to accelerate policy learning. The new “PicCoLOed” algorithm optimizes a policy by recursively repeating two steps: In the Prediction Step, the learner uses a model to predict the unseen future gradient and then applies the predicted estimate to update the policy; in the Correction Step, the learner runs the updated policy in the environment, receives the true gradient, and then corrects the policy using the gradient error. Unlike previous algorithms, PicCoLO corrects for the mistakes of using imperfect predicted gradients and hence does not suffer from model bias. The development of PicCoLO is made possible by a novel reduction from predictable online learning to adversarial online learning, which provides a systematic way to modify existing first-order algorithms to achieve the optimal regret with respect to predictable information. We show, in both theory and simulation, that the convergence rate of several first-order model-free algorithms can be improved by PicCoLO.

To perform complex tasks, robots must be able to interact with and manipulate their surroundings. One of the key challenges in accomplishing this is robust state estimation during physical interactions, where the state involves not only the robot and the object being manipulated, but also the state of the contact itself. In this work, within the context of planar pushing, we extend previous inference-based approaches to state estimation in several ways. We estimate the robot, object, and the contact state on multiple manipulation platforms configured with a vision-based articulated model tracker, and either a biomimetic tactile sensor or a force-torque sensor. We show how to fuse raw measurements from the tracker and tactile sensors to jointly estimate the trajectory of the kinematic states and the forces in the system via probabilistic inference on factor graphs, in both batch and incremental settings. We perform several benchmarks with our framework and show how performance is affected by incorporating various geometric and physics based constraints, occluding vision sensors, or injecting noise in tactile sensors. We also compare with prior work on multiple datasets and demonstrate that our approach can effectively optimize over multi-modal sensor data and reduce uncertainty to find better state estimates.

Current methods for estimating force from tactile sensor signals are either inaccurate analytic models or task-specific learned models. In this paper, we explore learning a robust model that maps tactile sensor signals to force. We specifically explore learning a mapping for the SynTouch BioTac sensor via neural networks. We propose a voxelized input feature layer for spatial signals and leverage information about the sensor surface to regularize the loss function. To learn a robust tactile force model that transfers across tasks, we generate ground truth data from three different sources: (1) the BioTac rigidly mounted to a force torque (FT) sensor, (2) a robot interacting with a ball rigidly attached to the same FT sensor, and (3) through force inference on a planar pushing task by formalizing the mechanics as a system of particles and optimizing over the object motion. A total of 140k samples were collected from the three sources. We achieve a median angular accuracy of 3.5 degrees in predicting force direction (66% improvement over the current state of the art) and a median magnitude accuracy of 0.06 N (93% improvement) on a test dataset. Additionally, we evaluate the learned force model in a force feedback grasp controller performing object lifting and gentle placement. Our results can be found on https: //sites.google.com/view/tactile-force.

Efficient planning in dynamic and uncertain environments is a fundamental challenge in robotics. In the context of trajectory optimization, the feasibility of paths can change as the environment evolves. Therefore, it can be beneficial to reason about multiple possible paths simultaneously. We build on prior work that considers graph-based trajectories to find solutions in multiple homotopy classes concurrently. Specifically, we extend this previous work to an online setting where the unreachable (in time) part of the graph is pruned and the remaining graph is reoptimized at every time step. As the robot moves within the graph on the path that is most promising, the pruning and reoptimization allows us to retain candidate paths that may become more viable in the future (a) as the environment changes, essentially enabling the robot to dynamically switch between numerous homotopy classes. We compare our approach against prior work without the homotopy switching capability and show improved performance across several metrics in simulation with a 2D robot in multiple dynamic environments under noisy measurements and execution.

Image alignment tasks require accurate pixel correspondences, which are usually recovered by matching local feature descriptors. Such descriptors are often derived using supervised learning on existing datasets with ground truth correspondences. However, the cost of creating such datasets is usually prohibitive. In this paper, we propose a new approach to align two images related by an unknown 2D homography where the local descriptor is learned from scratch from the images and the homography is estimated simultaneously. Our key insight is that a siamese convolutional neural network can be trained jointly while iteratively updating the homography parameters by optimizing a single loss function. Our method is currently weakly supervised because the input images need to be roughly aligned. We have used this method to align images of different modalities such as RGB and near-infra-red (NIR) without using any prior labeled data. Images automatically aligned by our method were then used to train descriptors that generalize to new images. We also evaluated our method on RGB images. On the HPatches benchmark, our method achieves comparable accuracy to deep local descriptors that were trained offline in a supervised setting.

Value aggregation is a general framework for solving imitation learning problems. Based on the idea of data aggregation, it generates a policy sequence by iteratively interleaving policy optimization and evaluation in an online learning setting. While the existence of a good policy in the policy sequence can be guaranteed non-asymptotically, little is known about the convergence of the sequence or the performance of the last policy. In this paper, we debunk the common belief that value aggregation always produces a convergent policy sequence with improving performance. Moreover, we identify a critical stability condition for convergence and provide a tight non-asymptotic bound on the performance of the last policy. These new theoretical insights let us stabilize problems with regularization, which removes the inconvenient process of identifying the best policy in the policy sequence in stochastic problems.

Hidden Quantum Markov Models (HQMMs) can be thought of as quantum probabilistic graphical models that can model sequential data. We extend previous work on HQMMs with three contributions: (1) we show how classical hidden Markov models (HMMs) can be simulated on a quantum circuit, (2) we reformulate HQMMs by relaxing the constraints for modeling HMMs on quantum circuits, and (3) we present a learning algorithm to estimate the parameters of an HQMM from data. While our algorithm requires further optimization to handle larger datasets, we are able to evaluate our algorithm using several synthetic datasets generated by valid HQMMs. We show that our algorithm learns HQMMs with the same number of hidden states and predictive accuracy as the HQMMs that generated the data, while HMMs learned with the Baum-Welch algorithm require more states to match the predictive accuracy.

Continuous-time trajectories are useful for reasoning about robot motion in a wide range of tasks. Sparse Gaussian processes (GPs) can be used as a non-parametric representation for trajectory distributions that enables fast trajectory optimization by sparse GP regression. However, most previous approaches that utilize sparse GPs for trajectory optimization are limited by the fact that the robot state is represented in vector space. In this paper, we first extend previous works to consider the state on general matrix Lie groups by applying a constant-velocity prior and defining locally linear GPs. Next, we discuss how sparse GPs on Lie groups provide a unified continuous-time framework for trajectory optimization for solving a number of robotics problems including state estimation and motion planning. Finally, we demonstrate and evaluate our approach on several different estimation and motion planning tasks with both synthetic and real-world experiments.

We consider the problems of learning forward models that map state to high-dimensional images and inverse models that map high-dimensional images to state in robotics. Specifically, we present a perceptual model for generating video frames from state with deep networks, and provide a framework for its use in tracking and prediction tasks. We show that our proposed model greatly outperforms standard deconvolutional methods and GANs for image generation, producing clear, photo-realistic images. We also develop a convolutional neural network model for state estimation and compare the result to an Extended Kalman Filter to estimate robot trajectories. We validate all models on a real robotic system.

A low-cost optical sensing method for improved measurement and control of soft pneumatic manipulator motion is presented. The core of a soft continuum robot is embedded with several optically-diffuse elastomer sensors which attenuate light depending on their strain mode and degree. The optical sensors measure local strains at the robot's axial center, and these strain data are combined with measured actuator chamber pressures to determine the pose of the robot under various gravitational and tip loading conditions. Regression analyses using neural networks (NNs) demonstrate that when the soft continuum robot's base orientation is fixed, the position of its end-effector can be estimated with 3.42 times more accuracy (71 % smaller root mean squared error) when using both optical sensor and pressure data (~2.44mm) than when using only pressure data (~8.3mm). When the robot's base orientation was varied, the combined optical sensor and pressure data provide position estimates which are as much as 37.8 times more accurate (~2.76mm) than pressure data alone (~104mm).

An understanding of the nature of objects could help robots to solve both high-level abstract tasks and improve performance at lower-level concrete tasks. Although deep learning has facilitated progress in image understanding, a robot's performance in problems like object recognition often depends on the angle from which the object is observed. Traditionally, robot sorting tasks rely on a fixed top-down view of an object. By changing its viewing angle, a robot can select a more semantically informative view leading to better performance for object recognition. In this paper, we introduce the problem of semantic view selection, which seeks to find good camera poses to gain semantic knowledge about an observed object. We propose a conceptual formulation of the problem, together with a solvable relaxation based on clustering. We then present a new image dataset consisting of around 10k images representing various views of 144 objects under different poses. Finally we use this dataset to propose a first solution to the problem by training a neural network to predict a “semantic score” from a top view image and camera pose. The views predicted to have higher scores are then shown to provide better clustering results than fixed top-down views.

Learning from Demonstration (LfD) is a popular approach to endowing robots with skills without having to program them by hand. Typically, LfD relies on human demonstrations in clutter-free environments. This prevents the demonstrations from being affected by irrelevant objects, whose influence can obfuscate the true intention of the human or the constraints of the desired skill. However, it is unrealistic to assume that the robot's environment can always be restructured to remove clutter when capturing human demonstrations. To contend with this problem, we develop an importance weighted batch and incremental skill learning approach, building on a recent inference-based technique for skill representation and reproduction. Our approach reduces unwanted environmental influences on the learned skill, while still capturing the salient human behavior. We provide both batch and incremental versions of our approach and validate our algorithms on a 7-DOF JACO2 manipulator with reaching and placing skills.

Recently, a novel class of Approximate Policy Iteration (API) algorithms have demonstrated impressive practical performance (e.g., ExIt from [1], AlphaGo-Zero from [2]). This new family of algorithms maintains, and alternately optimizes, two policies: a fast, reactive policy (e.g., a deep neural network) deployed at test time, and a slow, non-reactive policy (e.g., Tree Search), that can plan multiple steps ahead. The reactive policy is updated under supervision from the non-reactive policy, while the non-reactive policy is improved with guidance from the reactive policy. In this work we study this Dual Policy Iteration (DPI) strategy in an alternating optimization framework and provide a convergence analysis that extends existing API theory. We also develop a special instance of this framework which reduces the update of non-reactive policies to model-based optimal control using learned local models, and provides a theoretically sound way of unifying model-free and model-based RL approaches with unknown dynamics. We demonstrate the efficacy of our approach on various continuous control Markov Decision Processes.

We present foundations for using Model Predictive Control (MPC) as a differentiable policy class for reinforcement learning. This provides one way of leveraging and combining the advantages of model-free and model-based approaches. Specifically, we differentiate through MPC by using the KKT conditions of the convex approximation at a fixed point of the controller. Using this strategy, we are able to learn the cost and dynamics of a controller via end-to-end learning. Our experiments focus on imitation learning in the pendulum and cartpole domains, where we learn the cost and dynamics terms of an MPC policy class. We show that our MPC policies are significantly more data-efficient than a generic neural network and that our method is superior to traditional system identification in a setting where the expert is unrealizable.

Gaussian processes (GPs) provide a powerful non-parametric framework for reasoning over functions. Despite appealing theory, its superlinear computational and memory complexities have presented a long-standing challenge. State-of-the-art sparse variational inference methods trade modeling accuracy against complexity. However, the complexities of these methods still scale superlinearly in the number of basis functions, implying that that sparse GP methods are able to learn from large datasets only when a small model is used. Recently, a decoupled approach was proposed that removes the unnecessary coupling between the complexities of modeling the mean and the covariance functions of a GP. It achieves a linear complexity in the number of mean parameters, so an expressive posterior mean function can be modeled. While promising, this approach suffers from optimization difficulties due to ill-conditioning and non-convexity. In this work, we propose an alternative decoupled parametrization. It adopts an orthogonal basis in the mean function to model the residues that cannot be learned by the standard coupled approach. Therefore, our method extends, rather than replaces, the coupled approach to achieve strictly better performance. This construction admits a straightforward natural gradient update rule, so the structure of the information manifold that is lost during decoupling can be leveraged to speed up learning. Empirically, our algorithm demonstrates significantly faster convergence in multiple experiments.

Quantum Graphical Models (QGMs) generalize classical graphical models by adopting the formalism for reasoning about uncertainty from quantum mechanics. Unlike classical graphical models, QGMs represent uncertainty with density matrices in complex Hilbert spaces. Hilbert space embeddings (HSEs) also generalize Bayesian inference in Hilbert spaces. We investigate the link between QGMs and HSEs and show that the sum rule and Bayes rule for QGMs are equivalent to the kernel sum rule in HSEs and a special case of Nadaraya-Watson kernel regression, respectively. We show that these operations can be kernelized, and use these insights to propose a Hilbert Space Embedding of Hidden Quantum Markov Models (HSE-HQMM) to model dynamics. We present experimental results showing that HSE-HQMMs are competitive with state-of-the-art models like LSTMs and PSRNNs on several datasets, while also providing a nonparametric method for maintaining a probability distribution over continuous-valued features.

Many machine learning tasks, such as learning with invariance and policy evaluation in reinforcement learning, can be characterized as problems of learning from conditional distributions. In such problems, each sample x itself is associated with a conditional distribution $p(z|x)$ represented by samples $z_i_i=1^M$, and the goal is to learn a function f that links these conditional distributions to target values y. These problems become very challenging when we only have limited samples or in the extreme case only one sample from each conditional distribution. Commonly used approaches either assume that z is independent of x, or require an overwhelmingly large set of samples from each conditional distribution. To address these challenges, we propose a novel approach which employs a new min-max reformulation of the learning from conditional distribution problem. With such new reformulation, we only need to deal with the joint distribution p(z,x). We also design an efficient learning algorithm, Embedding-SGD, and establish theoretical sample complexity for such problems. Finally, our numerical experiments, on both synthetic and real-world datasets, show that the proposed approach can significantly improve over existing algorithms.

Sparse Spectrum Gaussian Processes (SSGPs) are a powerful tool for scaling Gaussian processes (GPs) to large datasets. Existing SSGP algorithms for regression assume deterministic inputs, precluding their use in many real-world robotics and engineering applications where accounting for input uncertainty is crucial. We address this problem by proposing two analytic moment-based approaches with closed-form expressions for SSGP regression with uncertain inputs. Our methods are more general and scalable than their standard GP counterparts, and are naturally applicable to multi-step prediction or uncertainty propagation. We show that efficient algorithms for Bayesian filtering and stochastic model predictive control can use these methods, and we evaluate our algorithms with comparative analyses and both real-world and simulated experiments.

Recently, researchers have demonstrated state-of-the-art performance on sequential prediction problems using deep neural networks and Reinforcement Learning (RL). For some of these problems, oracles that can demonstrate good performance may be available during training, but are not used by plain RL methods. To take advantage of this extra information, we propose AggreVaTeD, an extension of the Imitation Learning (IL) approach of Ross \& Bagnell (2014). AggreVaTeD allows us to use expressive differentiable policy representations such as deep networks, while leveraging training-time oracles to achieve faster and more accurate solutions with less training data. Specifically, we present two gradient procedures that can learn neural network policies for several problems, including a sequential prediction task and several high-dimensional robotics control problems. We also provide a comprehensive theoretical study of IL that demonstrates that we can expect up to exponentially-lower sample complexity for learning with AggreVaTeD than with plain RL algorithms. Our results and theory indicate that IL (and AggreVaTeD in particular) can be a more effective strategy for sequential prediction than plain RL.

Autonomous crop monitoring at high spatial and temporal resolution is a critical problem in precision agriculture. While Structure from Motion and Multi-View Stereo algorithms can finely reconstruct the 3D structure of a field with low-cost image sensors, these algorithms fail to capture the dynamic nature of continuously growing crops. In this paper we propose a 4D reconstruction approach to crop monitoring, which employs a spatio-temporal model of dynamic scenes that is useful for precision agriculture applications. Additionally, we provide a robust data association algorithm to address the problem of large appearance changes due to scenes being viewed from different angles at different points in time, which is critical to achieving 4D reconstruction. Finally, we collected a high-quality dataset with ground-truth statistics to evaluate the performance of our method. We demonstrate that our 4D reconstruction approach provides models that are qualitatively correct with respect to visual appearance and quantitatively accurate when measured against the ground truth geometric properties of the monitored crops.

Motion planning as trajectory optimization requires generating trajectories that minimize a desired objective function or performance metric. Finding a globally optimal solution is often intractable in practice: despite the existence of fast motion planning algorithms, most are prone to local minima, which may require re-solving the problem multiple times with different initializations. In this work we provide a novel motion planning algorithm, GPMP-GRAPH, that considers a graph-based initialization that simultaneously explores multiple homotopy classes, helping to contend with the local minima problem. Drawing on previous work to represent continuous-time trajectories as samples from a Gaussian process (GP) and formulating the motion planning problem as inference on a factor graph, we construct a graph of interconnected states such that each path through the graph is a valid trajectory and efficient inference can be performed on the collective factor graph. We perform a variety of benchmarks and show that our approach allows the evaluation of an exponential number of trajectories within a fraction of the computational time required to evaluate them one at a time, yielding a more thorough exploration of the solution space and a higher success rate.

The problem of optimal motion planing and control is fundamental in robotics. However, this problem is intractable for continuous-time stochastic systems in general and the solution is difficult to approximate if non-instantaneous nonlinear performance indices are present. In this work, we provide an efficient algorithm, PIPC (Probabilistic Inference for Planning and Control), that yields approximately optimal policies with arbitrary higher-order nonlinear performance indices. Using probabilistic inference and a Gaussian process representation of trajectories, PIPC exploits the underlying sparsity of the problem such that its complexity scales linearly in the number of nonlinear factors. We demonstrate the capabilities of our algorithm in a receding horizon setting with multiple systems in simulation.

We introduce an information theoretic model predictive control (MPC) algorithm capable of handling complex cost criteria and general nonlinear dynamics. The generality of the approach makes it possible to use multi-layer neural networks as dynamics models, which we incorporate into our MPC algorithm in order to solve model-based reinforcement learning tasks. We test the algorithm in simulation on a cart-pole swing up and quadrotor navigation task, as well as on actual hardware in an aggressive driving task. Empirical results demonstrate that the algorithm is capable of achieving a high level of performance and does so only utilizing data collected from the system.

Large-scale Gaussian process inference has long faced practical challenges due to time and space complexity that is superlinear in dataset size. While sparse variational Gaussian process models are capable of learning from large-scale data, standard strategies for sparsifying the model can prevent the approximation of complex functions. In this work, we propose a novel variational Gaussian process model that decouples the representation of mean and covariance functions in reproducing kernel Hilbert space. We show that this new parametrization generalizes previous models. Furthermore, it yields a variational inference problem that can be solved by stochastic gradient ascent with time and space complexity that is only linear in the number of mean function parameters, regardless of the choice of kernels, likelihoods, and inducing points. This strategy makes the adoption of large-scale expressive Gaussian process models possible. We run several experiments on regression tasks and show that this decoupled approach greatly outperforms previous sparse variational Gaussian process inference procedures.

We present a new model, Predictive State Recurrent Neural Networks (PSRNNs), for filtering and prediction in dynamical systems. PSRNNs draw on insights from both Recurrent Neural Networks (RNNs) and Predictive State Representations (PSRs), and inherit advantages from both types of models. Like many successful RNN architectures, PSRNNs use (potentially deeply composed) bilinear transfer functions to combine information from multiple sources. We show that such bilinear functions arise naturally from state updates in Bayes filters like PSRs, in which observations can be viewed as gating belief states. We also show that PSRNNs can be learned effectively by combining Backpropogation Through Time (BPTT) with an initialization derived from a statistically consistent learning algorithm for PSRs called two-stage regression (2SR). Finally, we show that PSRNNs can be factorized using tensor decomposition, reducing model size and suggesting interesting connections to existing multiplicative architectures such as LSTMs and GRUs. We apply PSRNNs to 4 datasets, and show that we outperform several popular alternative approaches to modeling dynamical systems in all cases.

Recurrent neural networks (RNNs) are a vital modeling technique that rely on internal states learned indirectly by optimization of a supervised, unsupervised, or reinforcement training loss. RNNs are used to model dynamic processes that are characterized by underlying latent states whose form is often unknown, precluding its analytic representation inside an RNN. In the Predictive-State Representation (PSR) literature, latent state processes are modeled by an internal state representation that directly models the distribution of future observations, and most recent work in this area has relied on explicitly representing and targeting sufficient statistics of this probability distribution. We seek to combine the advantages of RNNs and PSRs by augmenting existing state-of-the-art recurrent neural networks with Predictive-State Decoders (PSDs), which add supervision to the network's internal state representation to target predicting future observations. PSDs are simple to implement and easily incorporated into existing training pipelines via additional loss regularization. We demonstrate the effectiveness of PSDs with experimental results in three different domains: probabilistic filtering, Imitation Learning, and Reinforcement Learning. In each, our method improves statistical performance of state-of-the-art recurrent baselines and does so with fewer iterations and less data.

Instrumental variable regression (IVR) is a statistical technique utilized to recover unbiased estimators when there are errors in the independent variables. Estimator bias in learned time series models can yield poor performance in applications such as long-term prediction and filtering where the recursive use of the model results in the accumulation of propagated error. However, prior work addressed the IVR objective in the batch setting, where it is necessary to store the entire dataset in memory — an infeasible requirementin large dataset scenarios. In this work, we develop Online Instrumental Variable Regression (OIVR), an algorithm that is capable of updating the learned estimator with streaming data. We show that the online adaptation of IVR enjoys a no-regret performance guarantee with respect the original batchsetting by taking advantage of any no-regret online learning algorithm inside OIVR for the underlying update steps. We experimentally demonstrate the efficacy of our algorithm in combination with popular no-regret onlinealgorithms for the task of learning predictive dynamical system models and on a prototypical econometrics instrumental variable regression problem.

Recently, significant progress has been made developing kernel mean expressions for Bayesian inference. An important success in this domain is the nonparametric kernel Bayes’ filter (nKB-filter), which can be used for sequential inference in state space models. We expand upon this work by introducing a smoothing algorithm, the nonparametric kernel Bayes’ smoother (nKB-smoother) which relies on kernel Bayesian inference through the kernel sum rule and kernel Bayes’ rule. We derive the smoothing equations, analyze the computational cost, and show smoothing consistency. We summarize the algorithm, which is simple to implement, requiring only matrix multiplications and the output of the nKB-filter. Finally, we report experimental results that compare the nKB-smoother to previous parametric and nonparametric approaches to Bayesian filtering and smoothing. In the supplementary materials, we show that the combination of the nKB-filter and the nKB-smoother allows marginal kernel mean computation, which gives an alternative to kernel belief propagation.

Latent state space models are a fundamental and widely used tool for modeling dynamical systems. However, they are difficult to learn from data and learned models often lack performance guarantees on inference tasks such as filtering and prediction. In this work, we present the PREDICTIVE STATE INFERENCE MACHINE (PSIM), a data-driven method that considers the inference procedure on a dynamical system as a composition of predictors. The key idea is that rather than first learning a latent state space model, and then using the learned model for inference, PSIM directly learns predictors for inference in predictive state space. We provide theoretical guarantees for inference, in both realizable and agnostic settings, and showcase practical performance on a variety of simulated and real world robotics benchmarks.

Motion planning is a fundamental tool in robotics, used to generate collision-free, smooth, trajectories, while satisfying task-dependent constraints. In this paper, we present a novel approach to motion planning using Gaussian processes. In contrast to most existing trajectory optimization algorithms, which rely on a discrete state parameterization in practice, we represent the continuous-time trajectory as a sample from a Gaussian process (GP) generated by a linear time-varying stochastic differential equation. We then provide a gradient-based optimization technique that optimizes continuous-time trajectories with respect to a cost functional. By exploiting GP interpolation, we develop the Gaussian Process Motion Planner (GPMP), that finds optimal trajectories parameterized by a small number of states. We benchmark our algorithm against recent trajectory optimization algorithms by solving 7-DOF robotic arm planning problems in simulation and validate our approach on a real 7-DOF WAM arm.

Model-based trajectory optimization often fails to find a reference trajectory for under-actuated bipedal robots performing highly-dynamic, contact-rich tasks in the real world due to inaccurate physical models. In this paper, we propose a complete system that automatically designs a reference trajectory that succeeds on tasks in the real world with a very small number of real world experiments. We adopt existing system identification techniques and show that, with appropriate model parameterization and control optimization, an iterative system identification framework can be effective for designing reference trajectories. We focus on a set of tasks that leverage the momentum transfer strategy to rapidly change the whole-body from an initial configuration to a target configuration by generating large accelerations at the center of mass and switching contacts.

Recent work on scaling up Gaussian process regression (GPR) to large datasets has primarily focused on sparse GPR, which leverages a small set of basis functions to approximate the full Gaussian process during inference. However, the majority of these approaches are batch methods that operate on the entire training dataset at once, precluding the use of datasets that are streaming or too large to fit into memory. Although previous work has considered incrementally solving variational sparse GPR, most algorithms fail to update the basis functions and therefore perform suboptimally. We propose a novel incremental learning algorithm for variational sparse GPR based on stochastic mirror ascent of probability densities in reproducing kernel Hilbert space. This new formulation allows our algorithm to update basis functions online in accordance with the manifold structure of probability densities for fast convergence. We conduct several experiments and show that our proposed approach achieves better empirical performance in terms of prediction error than the recent state-of-the-art incremental solutions to variational sparse GPR.

In this paper, we attack the problem of learning a predictive model of a depth camera and manipulator directly from raw execution traces. While the problem of learning manipulator models from visual and proprioceptive data has been addressed before, existing techniques often rely on assumptions about the structure of the robot or tracked features in observation space. We make no such assumptions. Instead, we formulate the problem as that of learning a high-dimensional controlled stochastic process. We leverage recent work on nonparametric predictive state representations to learn a generative model of the depth camera and robotic arm from sequences of uninterpreted actions and observations. We perform several experiments in which we demonstrate that our learned model can accurately predict future depth camera observations in response to sequences of motor commands.

Functional gradient algorithms (e.g. CHOMP) have recently shown great promise for producing locally optimal motion for complex many degree-of-freedom robots. A key limitation of such algorithms is the difficulty in incorporating constraints and cost functions that explicitly depend on time. We present T-CHOMP, a functional gradient algorithm that overcomes this limitation by directly optimizing in space-time. We outline a framework for joint space-time optimization, derive an efficient trajectory-wide update for maintaining time monotonicity, and demonstrate the significance of T-CHOMP over CHOMP in several scenarios. By manipulating time, T-CHOMP produces lower-cost trajectories leading to behavior that is meaningfully different from CHOMP.

We present a novel spectral learning algorithm for simultaneous localization and mapping (SLAM) from range data with known correspondences. This algorithm is an instance of a general spectral system identification framework, from which it inherits several desirable properties, including statistical consistency and no local optima. Compared with popular batch optimization or multiple-hypothesis tracking (MHT) methods for range-only SLAM, our spectral approach offers guaranteed low computational requirements and good tracking performance. Compared with MHT and with popular extended Kalman filter (EKF) or extended information filter (EIF) approaches, our approach does not need to linearize a transition or measurement model. We provide a theoretical analysis of our method, including finite-sample error bounds. Finally, we demonstrate on a real-world robotic SLAM problem that our algorithm is not only theoretically justified, but works well in practice: in a comparison of multiple methods, the lowest errors come from a combination of our algorithm with batch optimization, but our method alone produces nearly as good a result at far lower computational cost.

Recently, a number of researchers have proposed spectral algorithms for learning models of dynamical systems — for example, Hidden Markov Models (HMMs), Partially Observable Markov Decision Processes (POMDPs), and Transformed Predictive State Representations (TPSRs). These algorithms are attractive since they are statistically consistent and not subject to local optima. However, they are batch methods: they need to store their entire training data set in memory at once and operate on it as a large matrix, and so they cannot scale to extremely large data sets (either many examples or many features per example). In turn, this restriction limits their ability to learn accurate models of complex systems. To overcome these limitations, we propose a new online spectral algorithm, which uses tricks such as incremental Singular Value Decomposition (SVD) and random projections to scale to much larger data sets and more complex systems than previous methods. We demonstrate the new method on an inertial measurement prediction task and a high-bandwidth video mapping task and we illustrate desirable behaviors such as "closing the loop," where the latent state representation changes suddenly as the learner recognizes that it has returned to a previously known place.

We propose a new approach to value function approximation which combines linear temporal difference reinforcement learning with subspace identification. In practical applications, reinforcement learning (RL) is complicated by the fact that state is either high-dimensional or partially observable. Therefore, RL methods are designed to work with features of state rather than state itself, and the success or failure of learning is often determined by the suitability of the selected features. By comparison, subspace identification (SSID) methods are designed to select a feature set which preserves as much information as possible about state. In this paper we connect the two approaches, looking at the problem of reinforcement learning with a large set of features, each of which may only be marginally useful for value function approximation. We introduce a new algorithm for this situation, called Predictive State Temporal Difference (PSTD) learning. As in SSID for predictive state representations, PSTD finds a linear compression operator that projects a large set of features down to a small set that preserves the maximum amount of predictive information. As in RL, PSTD then uses a Bellman recursion to estimate a value function. We discuss the connection between PSTD and prior approaches in RL and SSID. We prove that PSTD is statistically consistent, perform several experiments that illustrate its properties, and demonstrate its potential on a difficult optimal stopping problem.

Stability is a desirable characteristic for linear dynamical systems, but it is often ignored by algorithms that learn these systems from data. We propose a novel method for learning stable linear dynamical systems: we formulate an approxima- tion of the problem as a convex program, start with a solution to a relaxed version of the program, and incrementally add constraints to improve stability. Rather than continuing to generate constraints until we reach a feasible solution, we test stability at each step; because the convex program is only an approximation of the desired problem, this early stopping rule can yield a higher-quality solution. We apply our algorithm to the task of learning dynamic textures from image sequences as well as to modeling biosurveillance drug-sales data. The constraint generation approach leads to noticeable improvement in the quality of simulated sequences. We compare our method to those of Lacy and Bernstein [1, 2], with positive results in terms of accuracy, quality of simulated sequences, and efficiency.