Partial observability is a notorious challenge in reinforcement learning (RL), due to the need to learn complex, history-dependent policies. Recent empirical successes have used *privileged expert distillation* -- which leverages availability of latent state information during training (e.g., from a simulator) to learn and imitate the optimal latent, Markovian policy -- to disentangle the task of ''learning to see'' from ''learning to act''. While expert distillation is more computationally efficient than RL without latent state information, it also has well-documented failure modes. In this paper -- through a simple but instructive theoretical model called the *perturbed Block MDP*, and controlled experiments on challenging simulated locomotion tasks -- we investigate the algorithmic trade-off between privileged expert distillation and standard RL without privileged information. Our main findings are: **(1)** The trade-off empirically hinges on the *stochasticity* of the latent dynamics, as theoretically predicted by contrasting *approximate decodability* with *belief contraction* in the perturbed Block MDP; and **(2)** The optimal latent policy is not always the best latent policy to distill. Our results suggest new guidelines for effectively exploiting privileged information, potentially advancing the efficiency of policy learning across many practical partially observable domains.

We consider the hybrid reinforcement learning setting where the agent has access to both offline data and online interactive access. While RL research typically assumes offline data contains complete action, reward and transition information, datasets with only state information (also known as observation-only datasets) are more general, abundant and practical. This motivates our study of the hybrid RL with observation-only offline dataset framework. While the task of competing with the best policy “covered” by the offline data can be solved if a reset model of the environment is provided (i.e., one that can be reset to any state), we show evidence of hardness of competing when only given the weaker trace model (i.e., one can only reset to the initial states and must produce full traces through the environment), without further assumption of admissibility of the offline data. Under the admissibility assumptions– that the offline data could actually be produced by the policy class we consider– we propose the first algorithm in the trace model setting that provably matches the performance of algorithms that leverage a reset model. We also perform proof-of-concept experiments that suggest the effectiveness of our algorithm in practice.

The inverse reinforcement learning approach to imitation learning is a double-edged sword. On the one hand, it can enable learning from a smaller number of expert demonstrations with more robustness to error compounding than behavioral cloning approaches. On the other hand, it requires that the learner repeatedly solve a computationally expensive reinforcement learning (RL) problem. Often, much of this computation is wasted searching over policies very dissimilar to the expert's. In this work, we propose using *hybrid RL* -- training on a mixture of online and expert data -- to curtail unnecessary exploration. Intuitively, the expert data focuses the learner on good states during training, which reduces the amount of exploration required to compute a strong policy. Notably, such an approach doesn't need the ability to reset the learner to arbitrary states in the environment, a requirement of prior work in efficient inverse RL. More formally, we derive a reduction from inverse RL to *expert-competitive RL* (rather than globally optimal RL) that allows us to dramatically reduce interaction during the inner policy search loop while maintaining the benefits of the IRL approach. This allows us to derive both model-free and model-based hybrid inverse RL algorithms with strong policy performance guarantees. Empirically, we find that our approaches are significantly more sample efficient than standard inverse RL and several other baselines on a suite of continuous control tasks.

Learning from human preference data has emerged as the dominant paradigm for fine-tuning large language models (LLMs). The two most common families of techniques -- online reinforcement learning (RL) such as Proximal Policy Optimization (PPO) and offline contrastive methods such as Direct Preference Optimization (DPO) -- were positioned as equivalent in prior work due to the fact that both have to start from the same offline preference dataset. To further expand our theoretical understanding of the similarities and differences between online and offline techniques for preference fine-tuning, we conduct a rigorous analysis through the lens of *dataset coverage*, a concept that captures how the training data covers the test distribution and is widely used in RL. We prove that a global coverage condition is both necessary and sufficient for offline contrastive methods to converge to the optimal policy, but a weaker partial coverage condition suffices for online RL methods. This separation provides one explanation of why online RL methods can perform better than offline methods, especially when the offline preference data is not diverse enough. Finally, motivated by our preceding theoretical observations, we derive a hybrid preference optimization (HyPO) algorithm that uses offline data for contrastive-based preference optimization and online unlabeled data for KL regularization. Theoretically and empirically, we demonstrate that HyPO is more performant than its pure offline counterpart DPO, while still preserving its computation and memory efficiency.

While originally developed for continuous control problems, Proximal Policy Optimization (PPO) has emerged as the work-horse of a variety of reinforcement learning (RL) applications, including the fine-tuning of generative models. Unfortunately, PPO requires multiple heuristics to enable stable convergence (e.g. value networks, clipping), and is notorious for its sensitivity to the precise implementation of these components. In response, we take a step back and ask what a *minimalist* RL algorithm for the era of generative models would look like. We propose REBEL, an algorithm that cleanly reduces the problem of policy optimization to regressing the *relative reward* between two completions to a prompt in terms of the policy, enabling strikingly lightweight implementation. In theory, we prove that fundamental RL algorithms like Natural Policy Gradient can be seen as variants of REBEL, which allows us to match the strongest known theoretical guarantees in terms of convergence and sample complexity in the RL literature. REBEL can also cleanly incorporate offline data and be extended to handle the intransitive preferences we frequently see in practice. Empirically, we find that REBEL provides a unified approach to language modeling and image generation with stronger or similar performance as PPO and DPO, all while being simpler to implement and more computationally efficient than PPO. When fine-tuning Llama-3-8B-Instruct, REBEL achieves strong performance in AlpacaEval 2.0, MT-Bench, and Open LLM Leaderboard. Implementation of REBEL can be found at <https://github.com/ZhaolinGao/REBEL>, and models trained by REBEL can be found at <https://huggingface.co/Cornell-AGI>.

Inverse Reinforcement Learning (IRL) is a powerful set of techniques for imitation learning that aims to learn a reward function that rationalizes expert demonstrations. Unfortunately, traditional IRL methods suffer from a computational weakness: they require repeatedly solving a hard reinforcement learning (RL) problem as a subroutine. This is counter-intuitive from the viewpoint of reductions: we have reduced the easier problem of imitation learning to repeatedly solving the harder problem of RL. Another thread of work has proved that access to the side-information of the distribution of states where a strong policy spends time can dramatically reduce the sample and computational complexities of solving an RL problem. In this work, we demonstrate for the first time a more informed imitation learning reduction where we utilize the state distribution of the expert to alleviate the global exploration component of the RL subroutine, providing an exponential speedup in theory. In practice, we find that we are able to significantly speed up the prior art on continuous control tasks.

We propose a novel approach to addressing two fundamental challenges in Model-based Reinforcement Learning (MBRL): the computational expense of repeatedly finding a good policy in the learned model, and the objective mismatch between model fitting and policy computation. Our "lazy" method leverages a novel unified objective, Performance Difference via Advantage in Model, to capture the performance difference between the learned policy and expert policy under the true dynamics. This objective demonstrates that optimizing the expected policy advantage in the learned model under an exploration distribution is sufficient for policy computation, resulting in a significant boost in computational efficiency compared to traditional planning methods. Additionally, the unified objective uses a value moment matching term for model fitting, which is aligned with the model’s usage during policy computation. We present two no-regret algorithms to optimize the proposed objective, and demonstrate their statistical and computational gains compared to existing MBRL methods through simulated benchmarks.

We develop algorithms for imitation learning from policy data that was corrupted by temporally correlated noise in expert actions. When noise affects multiple timesteps of recorded data, it can manifest as spurious correlations between states and actions that a learner might latch on to, leading to poor policy performance. To break up these spurious correlations, we apply modern variants of the instrumental variable regression (IVR) technique of econometrics, enabling us to recover the underlying policy without requiring access to an interactive expert. In particular, we present two techniques, one of a generative-modeling flavor (DoubIL) that can utilize access to a simulator, and one of a game-theoretic flavor (ResiduIL) that can be run entirely offline. We find both of our algorithms compare favorably to behavioral cloning on simulated control tasks.

We consider imitation learning problems where the learner's ability to mimic the expert increases throughout the course of an episode as more information is revealed. One example of this is when the expert has access to privileged information: while the learner might not be able to accurately reproduce expert behavior early on in an episode, by considering the entire history of states and actions, they might be able to eventually identify the hidden context and act as the expert would. We prove that on-policy imitation learning algorithms (with or without access to a queryable expert) are better equipped to handle these sorts of asymptotically realizable problems than off-policy methods. This is because on-policy algorithms provably learn to recover from their initially suboptimal actions, while off-policy methods treat their suboptimal past actions as though they came from the expert. This often manifests as a latching behavior: a naive repetition of past actions. We conduct experiments in a toy bandit domain that show that there exist sharp phase transitions of whether off-policy approaches are able to match expert performance asymptotically, in contrast to the uniformly good performance of on-policy approaches. We demonstrate that on several continuous control tasks, on-policy approaches are able to use history to identify the context while off-policy approaches actually perform worse when given access to history.

Online imitation learning is the problem of how best to mimic expert demonstrations, given access to the environment or an accurate simulator. Prior work has shown that in the \textit{infinite} sample regime, exact moment matching achieves value equivalence to the expert policy. However, in the \textit{finite} sample regime, even if one has no optimization error, empirical variance can lead to a performance gap that scales with $H^2 / N_{\text{exp}}$ for behavioral cloning and $H / N_{\text{exp}}$ for online moment matching, where $H$ is the horizon and $N_{\text{exp}}$ is the size of the expert dataset. We introduce the technique of ``replay estimation'' to reduce this empirical variance: by repeatedly executing cached expert actions in a stochastic simulator, we compute a smoother expert visitation distribution estimate to match. In the presence of general function approximation, we prove a meta theorem reducing the performance gap of our approach to the \textit{parameter estimation error} for offline classification (i.e. learning the expert policy). In the tabular setting or with linear function approximation, our meta theorem shows that the performance gap incurred by our approach achieves the optimal $\widetilde{O} \left( \min( H^{3/2} / N_{\text{exp}}, H / \sqrt{N_{\text{exp}}} \right)$ dependency, under significantly weaker assumptions compared to prior work. We implement multiple instantiations of our approach on several continuous control tasks and find that we are able to significantly improve policy performance across a variety of dataset sizes.

Given access to accurate dynamical models, modern planning approaches are effective in computing feasible and optimal plans for repetitive robotic tasks. However, it is difficult to model the true dynamics of the real world before execution, especially for tasks requiring interactions with objects whose parameters are unknown. A recent planning approach, CMAX, tackles this problem by adapting the planner online during execution to bias the resulting plans away from inaccurately modeled regions. CMAX, while being provably guaranteed to reach the goal, requires strong assumptions on the accuracy of the model used for planning and fails to improve the quality of the solution over repetitions of the same task. In this paper we propose CMAX++, an approach that leverages real-world experience to improve the quality of resulting plans over successive repetitions of a robotic task. CMAX++ achieves this by integrating model-free learning using acquired experience with model-based planning using the potentially inaccurate model. We provide provable guarantees on the completeness and asymptotic convergence of CMAX++ to the optimal path cost as the number of repetitions increases. CMAX++ is also shown to outperform baselines in simulated robotic tasks including 3D mobile robot navigation where the track friction is incorrectly modeled, and a 7D pick-and-place task where the mass of the object is unknown leading to discrepancy between true and modeled dynamics.

We provide a unifying view of a large family of previous imitation learning algorithms through the lens of moment matching. At its core, our classification scheme is based on whether the learner attempts to match (1) reward or (2) action-value moments of the expert’s behavior, with each option leading to differing algorithmic approaches. By considering adversarially chosen divergences between learner and expert behavior, we are able to derive bounds on policy performance that apply for all algorithms in each of these classes, the first to our knowledge. We also introduce the notion of moment recoverability, implicit in many previous analyses of imitation learning, which allows us to cleanly delineate how well each algorithmic family is able to mitigate compounding errors. We derive three novel algorithm templates (AdVIL, AdRIL, and DAeQuIL) with strong guarantees, simple implementation, and competitive empirical performance.

This work considers the trade-off between accuracy and testtime computational cost of deep neural networks (DNNs) via anytime predictions from auxiliary predictions. Specifically, we optimize auxiliary losses jointly in an adaptive weighted sum, where the weights are inversely proportional to average of each loss. Intuitively, this balances the losses to have the same scale. We demonstrate theoretical considerations that motivate this approach from multiple viewpoints, including connecting it to optimizing the geometric mean of the expectation of each loss, an objective that ignores the scale of losses. Experimentally, the adaptive weights induce more competitive anytime predictions on multiple recognition data-sets and models than non-adaptive approaches including weighing all losses equally. In particular, anytime neural networks (ANNs) can achieve the same accuracy faster using adaptive weights on a small network than using static constant weights on a large one. For problems with high performance saturation, we also show a sequence of exponentially deepening ANNs can achieve near-optimal anytime results at any budget, at the cost of a const fraction of extra computation.

Black-box optimizers that explore in parameter space have often been shown to outperform more sophisticated action space exploration methods developed specifically for the reinforcement learning problem. We examine these black-box methods closely to identify situations in which they are worse than action space exploration methods and those in which they are superior. Through simple theoretical analyses, we prove that complexity of exploration in parameter space depends on the dimensionality of parameter space, while complexity of exploration in action space depends on both the dimensionality of action space and horizon length. This is also demonstrated empirically by comparing simple exploration methods on several model problems, including Contextual Bandit, Linear Regression and Reinforcement Learning in continuous control.

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.

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.

Boosting is a popular ensemble algorithm that generates more powerful learners by linearly combining base models from a simpler hypothesis class. In this work, we investigate the problem of adapting batch gradient boosting for minimizing convex loss functions to online setting where the loss at each iteration is i.i.d sampled from an unknown distribution. To generalize from batch to online, we first introduce the definition of online weak learning edge with which for strongly convex and smooth loss functions, we present an algorithm, Streaming Gradient Boosting (SGB) with exponential shrinkage guarantees in the number of weak learners. We further present an adaptation of SGB to optimize non-smooth loss functions, for which we derive a $O(\ln(N)/N)$ convergence rate. We also show that our analysis can extend to adversarial online learning setting under a stronger assumption that the online weak learning edge will hold in adversarial setting. We finally demonstrate experimental results showing that in practice our algorithms can achieve competitive results as classic gradient boosting while using less computation.

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.

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.

In shared autonomy, user input and robot autonomy are combined to control a robot to achieve a goal. One often used strategy considers the user and autonomy as independent decision makers, with the system blending these decisions. However, this independence leads to suboptimal, and often frustrating, behavior. Instead, we propose a system that explicitly models the interplay between the user and assistance. Our approach centers around the idea of learning how users respond to assistance. We then propose a cost minimization framework for assisting while utilizing this learned model.

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.

We propose a polynomial force-motion model for planar sliding. The set of generalized friction loads is the 1-sublevel set of a polynomial whose gradient directions correspond to generalized velocities. Additionally, the polynomial is confined to be convex even-degree homogeneous in order to obey the maximum work inequality, symmetry, shape invariance in scale, and fast invertibility. We present a simple and statistically-efficient model identification procedure using a sum-of-squares convex relaxation. Simulation and robotic experiments validate the accuracy and efficiency of our approach. We also show practical applications of our model including stable pushing of objects and free sliding dynamic simulations.

As robots aspire for long-term autonomous operations in complex dynamic environments, the ability to reliably take mission-critical decisions in ambiguous situations becomes critical. This motivates the need to build systems that have situational awareness to assess how qualified they are at that moment to make a decision. We call this self-evaluating capability as introspection. In this paper, we take a small step in this direction and propose a generic framework for introspective behavior in perception systems. Our goal is to learn a model to reliably predict failures in a given system, with respect to a task, directly from input sensor data. We present this in the context of vision-based autonomous MAV flight in outdoor natural environments, and show that it effectively handles uncertain situations.

We present a fully autonomous robotic system for grasping objects in dense clutter. The objects are unknown and have arbitrary shapes. Therefore, we cannot rely on prior models. Instead, the robot learns online, from scratch, to manipulate the objects by trial and error. Grasping objects in clutter is significantly harder than grasping isolated objects, because the robot needs to push and move objects around in order to create sufficient space for the fingers. These pre-grasping actions do not have an immediate utility, and may result in unnecessary delays. The utility of a pre-grasping action can be measured only by looking at the complete chain of consecutive actions and effects. This is a sequential decision-making problem that can be cast in the reinforcement learning framework. We solve this problem by learning the stochastic transitions between the observed states, using nonparametric density estimation. The learned transition function is used only for re-calculating the values of the executed actions in the observed states, with different policies. Values of new state-actions are obtained by regressing the values of the executed actions. The state of the system at a given time is a depth (3D) image of the scene. We use spectral clustering for detecting the different objects in the image. The performance of our system is assessed on a robot with real-world objects.

How should we gather information to make effective decisions? A classical answer to this fundamental problem is given by the decision-theoretic value of information. Unfortunately, optimizing this objective is intractable, and myopic (greedy) approximations are known to perform poorly. In this paper, we introduce DiRECt, an efficient yet near-optimal algorithm for nonmyopically optimizing value of information. Crucially, DiRECt uses a novel surrogate objective that is: (1) aligned with the value of information problem (2) efficient to evaluate and (3) adaptive submodular. This latter property enables us to utilize an efficient greedy optimization while providing strong approximation guarantees. We demonstrate the utility of our approach on four diverse case-studies: touch-based robotic localization, comparison-based preference learning, wild-life conservation management, and preference elicitation in behavioral economics. In the first application, we demonstrate DiRECt in closed-loop on an actual robotic platform.

Maximum entropy inverse optimal control (MaxEnt IOC) is an effective means of discovering the underlying cost function of demonstrated human activity and can be used to predict human behavior over low-dimensional state spaces (i.e., forecasting of 2D trajectories). To enable inference in very large state spaces, we introduce an approximate MaxEnt IOC procedure to address the fundamental computational bottleneck stemming from calculating the partition function via dynamic programming. Approximate MaxEnt IOC is based on two components: approximate dynamic programming and Monte Carlo sampling. We analyze this approximation approach and provide a finite-sample error upper bound on its excess loss. We validate the proposed method in the context of analyzing dual-agent interactions from video, where we use approximate MaxEnt IOC to simulate mental images of a single agents body pose sequence (a high-dimensional image space). We experiment with sequences image data taken from RGB and RGBD data and show that it is possible to learn cost functions that lead to accurate predictions in high-dimensional problems that were previously intractable.

Most typical statistical and machine learning approaches to time series modeling optimize a single-step prediction error. In multiple-step simulation, the learned model is iteratively applied, feeding through the previous output as its new input. Any such predictor however, inevitably introduces errors, and these compounding errors change the input distribution for future prediction steps, breaking the train-test i.i.d assumption common in supervised learning. We present an approach that reuses training data to make a no-regret learner robust to errors made during multi-step prediction. Our insight is to formulate the problem as imitation learning; the training data serves as a "demonstrator" by providing corrections for the errors made during multi-step prediction. By this reduction of multi-step time series prediction to imitation learning, we establish theoretically a strong performance guarantee on the relation between training error and the multi-step prediction error. We present experimental results of our method, DaD, and show significant improvement over the traditional approach in two notably different domains, dynamic system modeling and video texture prediction.

We propose a novel online learning method for minimizing regret in large extensive-form games. The approach learns a function approximator online to estimate the regret for choosing a particular action. A no-regret algorithm uses these estimates in place of the true regrets to define a sequence of policies. We prove the approach sound by providing a bound relating the quality of the function approximation and regret of the algorithm. A corollary being that the method is guaranteed to converge to a Nash equilibrium in self-play so long as the regrets are ultimately realizable by the function approximator. Our technique can be understood as a principled generalization of existing work onabstraction in large games; in our work, both the abstraction as well as the equilibrium are learned during self-play. We demonstrate empirically the method achieves higher quality strategies than state-of-the-art abstraction techniques given the same resources.

We present a simple approach for producing a small number of structured visual outputs which have high recall, for a variety of tasks including monocular pose estimation and semantic scene segmentation. Current state-of-the-art approaches learn a single model and modify inference procedures to produce a small number of diverse predictions. We take the alternate route of modifying the learning procedure to directly optimize for good, high recall sequences of structured-output predictors. Our approach introduces no new parameters, naturally learns diverse predictions and is not tied to any specific structured learning or inference procedure. We leverage recent advances in the contextual submodular maximization literature to learn a sequence of predictors and empirically demonstrate the simplicity and performance of our approach on multiple challenging vision tasks including achieving state-of-the-art results on multiple predictions for monocular pose-estimation and image foreground/background segmentation.

We formalize the problem of adapting a demonstrated trajectory to a new start and goal configuration as an optimization problem over a Hilbert space of trajectories: minimize the distance between the demonstration and the new trajectory subject to the new end point constraints. We show that the commonly used version of Dynamic Movement Primitives (DMPs) implement this minimization in the way they adapt demonstrations, for a particular choice of the Hilbert space norm. The generalization to arbitrary norms enables the robot to select a more appropriate norm for the task, as well as learn how to adapt the demonstration from the user. Our experiments show that this can significantly improve the robot's ability to accurately generalize the demonstration.

We consider detecting objects in an image by iteratively selecting from a set of arbitrarily shaped candidate regions. Our generic approach, which we term visual chunking, reasons about the locations of multiple object instances in an image while expressively describing object boundaries. We design an optimization criterion for measuring the performance of a list of such detections as a natural extension to a common per-instance metric. We present an efficient algorithm with provable performance for building a high-quality list of detections from any candidate set of region-based proposals. We also develop a simple class-specific algorithm to generate a candidate region instance in near-linear time in the number of low-level superpixels that outperforms other region generating methods. In order to make predictions on novel images at testing time without access to ground truth, we develop learning approaches to emulate these algorithms' behaviors. We demonstrate that our new approach outperforms sophisticated baselines on benchmark datasets.

Manipulating natural objects of irregular shapes, such as rocks, is an essential capability of robots operating in outdoor environments. Physics-based simulators are commonly used to plan stable grasps for man-made objects. However, planning is an expensive process that is based on simulating hand and object trajectories in different configurations, and evaluating the outcome of each trajectory. This problem is particularly concerning when the objects are irregular or cluttered, because the space of feasible grasps is significantly smaller, and more configurations need to be evaluated before finding a good one. In this paper, we first present a learning technique for fast detection of an initial set of potentially stable grasps in a cluttered scene. The best detected grasps are further optimized by fine-tuning the configuration of the hand in simulation. To reduce the computational burden of this last operation, we model the outcomes of the grasps as a Gaussian Process, and use an entropy-search method in order to focus the optimization on regions where the best grasp is most likely to be. This approach is tested on the task of clearing piles of real, unknown, rock debris with an autonomous robot. Empirical results show a clear advantage of the proposed approach when the time window for decision is short.

How should we gather information to make effective decisions? We address Bayesian active learning and experimental design problems, where we sequentially select tests to reduce uncertainty about a set of hypotheses. Instead of minimizing uncertainty per se, we consider a set of overlapping decision regions of these hypotheses. Our goal is to drive uncertainty into a single decision region as quickly as possible. We identify necessary and sufficient conditions for correctly identifying a decision region that contains all hypotheses consistent with observations. We develop a novel Hyperedge Cutting (HEC) algorithm for this problem, and prove that is competitive with the intractable optimal policy. Our efficient implementation of the algorithm relies on computing subsets of the complete homogeneous symmetric polynomials. Finally, we demonstrate its effectiveness on two practical applications: approximate comparison-based learning and active localization using a robot manipulator.

Many prediction domains, such as ad placement, recommendation, trajectory prediction, and document summarization, require predicting a set or list of options. Such lists are often evaluated using submodular reward functions that measure both quality and diversity. We propose a simple, efficient, and provably near-optimal approach to optimizing such prediction problems based on no-regret learning. Our method leverages a surprising result from online submodular optimization: a single no-regret online learner can compete with an optimal sequence of predictions. Compared to previous work, which either learn a sequence of classifiers or rely on stronger assumptions such as realizability, we ensure both data-efficiency as well as performance guarantees in the fully agnostic setting. Experiments validate the efficiency and applicability of the approach on a wide range of problems including manipulator trajectory optimization, news recommendation and document summarization.

Rich scene understanding from 3-D point clouds is a challenging task that requires contextual reasoning, which is typically computationally expensive. The task is further complicated when we expect the scene analysis algorithm to also efficiently handle data that is continuously streamed from a sensor on a mobile robot. Hence, we are typically forced to make a choice between 1) using a precise representation of the scene at the cost of speed, or 2) making fast, though inaccurate, approximations at the cost of increased misclassifications. In this work, we demonstrate that we can achieve the best of both worlds by using an efficient and simple representation of the scene in conjunction with recent developments in structured prediction in order to obtain both efficient and state-of-the-art classifications. Furthermore, this efficient scene representation naturally handles streaming data and provides a 300% to 500% speedup over more precise representations.

Many robotic systems deal with uncertainty by performing a sequence of information gathering actions. In this work, we focus on the problem of efficiently constructing such a sequence by drawing an explicit connection to submodularity. Ideally, we would like a method that finds the optimal sequence, taking the minimum amount of time while providing sufficient information. Finding this sequence, however, is generally intractable. As a result, many well-established methods select actions greedily. Surprisingly, this often performs well. Our work first explains this high performance - we note a commonly used metric, reduction of Shannon entropy, is submodular under certain assumptions, rendering the greedy solution comparable to the optimal plan in the offline setting. However, reacting online to observations can increase performance. Recently developed notions of adaptive submodularity provide guarantees for a greedy algorithm in this online setting. In this work, we develop new methods based on adaptive submodularity for selecting a sequence of information gathering actions online. In addition to providing guarantees, we can capitalize on submodularity to attain additional computational speedups. We demonstrate the effectiveness of these methods in simulation and on a robot.

We address the problem of clearing a pile of unknown objects using an autonomous interactive perception approach. Our robot hypothesizes the boundaries of objects in a pile of unknown objects (object segmentation) and verifies its hypotheses (object detection) using deliberate interactions. To guarantee the safety of the robot and the environment, we use compliant motion primitives for poking and grasping. Every verified segmentation hypothesis can be used to parameterize a compliant controller for manipulation or grasping. The robot alternates between poking actions to verify its segmentation and grasping actions to remove objects from the pile. We demonstrate our method with a robotic manipulator. We evaluate our approach with real-world experiments of clearing cluttered scenes composed of unknown objects.

We present an interactive perceptual skill for segmenting, tracking, and modeling the kinematic structure of 3D articulated objects. This skill is a prerequisite for general manipulation in unstructured environments. Robot-environment interactions are used to move an unknown object, creating a perceptual signal that reveals the kinematic properties of the object. The resulting perceptual information can then inform and facilitate further manipulation. The algorithm is computationally efficient, handles partial occlusions, and depends on little object motion; it only requires sufficient texture for visual feature tracking. We conducted experiments with everyday objects on a robotic manipulation platform equipped with an RGB-D sensor. The results demonstrate the robustness of the proposed method to lighting conditions, object appearance, size, structure, and configuration.

We address the problem of image-based scene analysis from streaming video, as would be seen from a moving platform, in order to efficiently generate spatially and temporally consistent predictions of semantic categories over time. In contrast to previous techniques which typically address this problem in batch and/or through graphical models, we demonstrate that by learning visual similarities between pixels across frames, a simple filtering algorithfiltering algorithmm is able to achieve high performance predictions in an efficient and online/causal manner. Our technique is a meta-algorithm that can be efficiently wrapped around any scene analysis technique that produces a per-pixel semantic category distribution. We validate our approach over three different scene analysis techniques on three different datasets that contain different semantic object categories. Our experiments demonstrate that our approach is very efficient in practice and substantially improves the consistency of the predictions over time.

Autonomous navigation for large Unmanned Aerial Vehicles (UAVs) is fairly straight-forward, as expensive sensors and monitoring devices can be employed. In contrast, obstacle avoidance remains a challenging task for Micro Aerial Vehicles (MAVs) which operate at low altitude in cluttered environments. Unlike large vehicles, MAVs can only carry very light sensors, such as cameras, making autonomous navigation through obstacles much more challenging. In this paper, we describe a system that navigates a small quadrotor helicopter autonomously at low altitude through natural forest environments. Using only a single cheap camera to perceive the environment, we are able to maintain a constant velocity of up to 1.5m/s. Given a small set of human pilot demonstrations, we use recent state-of-the-art imitation learning techniques to train a controller that can avoid trees by adapting the MAVs heading. We demonstrate the performance of our system in a more controlled environment indoors, and in real natural forest environments outdoors.

The principle of maximum entropy provides a powerful framework for estimating joint, conditional, and marginal probability distributions. However, there are many important distributions with elements of interaction and feedback where its applicability has not been established. This paper presents the principle of maximum causal entropy-an approach based on directed information theory for estimating an unknown process based on its interactions with a known process. We demonstrate the breadth of the approach using two applications: a predictive solution for inverse optimal control in decision processes and computing equilibrium strategies in sequential games.

A popular approach to high dimensional control problems in robotics uses a library of candidate “maneuvers” or “trajectories”. The library is either evaluated on a fixed number of candidate choices at runtime (e.g. path set selection for planning) or by iterating through a sequence of feasible choices until success is achieved (e.g. grasp selection). The performance of the library relies heavily on the content and order of the sequence of candidates. We propose a provably efficient method to optimize such libraries, leveraging recent advances in optimizing submodular functions of sequences. This approach is demonstrated on two important problems: mobile robot navigation and manipulator grasp set selection. In the first case, performance can be improved by choosing a subset of candidates which optimizes the metric under consideration (cost of traversal). In the second case, performance can be optimized by minimizing the depth in the list that is searched before a successful candidate is found. Our method can be used in both on-line and batch settings with provable performance guarantees, and can be run in an anytime manner to handle real-time constraints.

Building robust and reliable autonomous navigation systems that generalize across environments and operating scenarios remains a core challenge in robotics. Machine learning has proven a significant aid in this task; in recent years learning from demonstration has become especially popular, leading to improved systems while requiring less expert tuning and interaction. However, these approaches still place a burden on the expert, specifically to choose the best demonstrations to provide. This work proposes two approaches for active learning from demonstration, in which the learning system requests specific demonstrations from the expert. The approaches identify examples for which expert demonstration is predicted to provide useful information on concepts which are either novel or uncertain to the current system. Experimental results demonstrate both improved generalization performance and reduced expert interaction when using these approaches.

Search based planners such as A* and Dijkstra's algorithm are proven methods for guiding today's robotic systems. Although such planners are typically based upon a coarse approximation of reality, they are nonetheless valuable due to their ability to reason about the future, and to generalize to previously unseen scenarios. However, encoding the desired behavior of a system into the underlying cost function used by the planner can be a tedious and error-prone task. We introduce Reinforcement Planning, which extends gradient based reinforcement learning algorithms to automatically learn useful surrogate cost functions for optimal planners. Reinforcement Planning presents several advantages over other learning approaches to planning in that it is not limited by the expertise of a human demonstrator, and that it acknowledges the domain of the planner is a simplified model of the world. We demonstrate the effectiveness of our method in learning to solve a noisy physical simulation of the well-known “marble maze” toy.

We describe the software components of a robotics system designed to autonomously grasp objects and perform dexterous manipulation tasks with only high-level supervision. The system is centered on the tight integration of several core functionalities, including perception, planning and control, with the logical structuring of tasks driven by a Behavior Tree architecture. The advantage of the implementation is to reduce the execution time while integrating advanced algorithms for autonomous manipulation. We describe our approach to 3-D perception, real-time planning, force compliant motions, and audio processing. Performance results for object grasping and complex manipulation tasks of in-house tests and of an independent evaluation team are presented.

Numerous interaction techniques have been developed that make "virtual" pointing at targets in graphical user interfaces easier than analogous physical pointing tasks by invoking target-based interface modifications. These pointing facilitation techniques crucially depend on methods for estimating the relevance of potential targets. Unfortunately, many of the simple methods employed to date are inaccurate in common settings with many selectable targets in close proximity. In this paper, we bring recent advances in statistical machine learning to bear on this underlying target relevance estimation problem. By framing past target-driven pointing trajectories as approximate solutions to well-studied control problems, we learn the probabilistic dynamics of pointing trajectories that enable more accurate predictions of intended targets.

The application of the maximum entropy principle to sequence modeling has been popularized by methods such as Conditional Random Fields (CRFs). However, these approaches are generally limited to modeling paths in discrete spaces of low dimensionality. We consider the problem of modeling distributions over paths in continuous spaces of high dimensionality---a problem for which inference is generally intractable. Our main contribution is to show that maximum entropy modeling of high-dimensional, continuous paths is tractable as long as the constrained features possess a certain kind of low dimensional structure. In this case, we show that the associated {\em partition function} is symmetric and that this symmetry can be exploited to compute the partition function efficiently in a compressed form. Empirical results are given showing an application of our method to maximum entropy modeling of high dimensional human motion capture data.

Nearly every structured prediction problem in computer vision requires approximate inference due to large and complex dependencies among output labels. While graphical models provide a clean separation between modeling and inference, learning these models with approximate inference is not well understood. Furthermore, even if a good model is learned, predictions are often inaccurate due to approximations. In this work, instead of performing inference over a graphical model, we instead consider the inference procedure as a composition of predictors. Specifically, we focus on message-passing algorithms, such as Belief Propagation, and show how they can be viewed as procedures that sequentially predict label distributions at each node over a graph. Given labeled graphs, we can then train the sequence of predictors to output the correct labeling s. The result no longer corresponds to a graphical model but simply defines an inference procedure, with strong theoretical properties, that can be used to classify new graphs. We demonstrate the scalability and efficacy of our approach on 3D point cloud classification and 3D surface estimation from single images.

The high cost of damaging an expensive robot or injuring people or equipment in its environment make even rare failures unacceptable in many mobile robot applications. Often the objects that pose the highest risk for a mobile robot are those that were not present throughout previous successful traversals of an environment. Change detection, a closely related problem to novelty detection, is therefore of high importance to many mobile robotic applications that require a robot to operate repeatedly in the same environment. We present a novel algorithm for performing online change detection based on a previously developed robust online novelty detection system that uses a learned lower-dimensional representation of the feature space to perform measures of similarity. We then further improve this change detection system by incorporating online scene segmentation to better utilize contextual information in the environment. We validate these approaches through extensive experiments onboard a large outdoor mobile robot. Our results show that our approaches are robust to noisy sensor data and moderate registration errors and maintain their performance across diverse natural environments and conditions.

We address the problem of understanding scenes from 3-D laser scans via per-point assignment of semantic labels. In order to mitigate the difficulties of using a graphical model for modeling the contextual relationships among the 3-D points, we instead propose a multi-stage inference procedure to capture these relationships. More specifically, we train this procedure to use point cloud statistics and learn relational information (e.g., tree-trunks are below vegetation) over fine (point-wise) and coarse (region-wise) scales. We evaluate our approach on three different datasets, that were obtained from different sensors, and demonstrate improved performance.

Novelty detection is often treated as a one-class classification problem: how to segment a data set of examples from everything else that would be considered novel or abnormal. Almost all existing novelty detection techniques, however, suffer from diminished performance when the number of less relevant, redundant or noisy features increases, as often the case with high-dimensional feature spaces. Many of these algorithms are also not suited for online use, a trait that is highly desirable for many robotic applications. We present a novelty detection algorithm that is able to address this sensitivity to high feature dimensionality by utilizing prior class information within the training set. Additionally, our anytime algorithm is well suited for online use when a constantly adjusting environmental model is beneficial. We apply this algorithm to online detection of novel perception system input on an outdoor mobile robot and argue such abilities could be key in increasing the real-world applications and impact of mobile robotics1.

We present a novel approach to legged locomotion over rough terrain that is thoroughly rooted in optimization. This approach relies on a hierarchy of fast, anytime algorithms to plan a set of footholds, along with the dynamic body motions required to execute them. Components within the planning framework coordinate to exchange plans, cost-to-go estimates, and “certificates” that ensure the output of an abstract high-level planner can be realized by deeper layers of the hierarchy. The burden of careful engineering of cost functions to achieve desired performance is substantially mitigated by a simple inverse optimal control technique. Robustness is achieved by real-time re-planning of the full trajectory, augmented by reflexes and feedback control. We demonstrate the successful application of our approach in guiding the LittleDog quadruped robot over a variety of rough terrains.

We address the problem of label assignment in computer vision: given a novel 3D or 2D scene, we wish to assign a unique label to every site (voxel, pixel, superpixel, etc.). To this end, the Markov Random Field framework has proven to be a model of choice as it uses contextual information to yield improved classification results over locally independent classifiers. In this work we adapt a functional gradient approach for learning high-dimensional parameters of random fields in order to perform discrete, multi-label classification. With this approach we can learn robust models involving high-order interactions better than the previously used learning method. We validate the approach in the context of point cloud classification and improve the state of the art. In addition, we successfully demonstrate the generality of the approach on the challenging vision problem of recovering 3-D geometric surfaces from images.

We present GATMO (Generalized Approach to Tracking Movable Objects), a system for localization and mapping that incorporates the dynamic nature of the environment while maintaining semantic labels. Objects in the environment are broken down into multiple mobility levels, from static (walls) to highly mobile (people), by maintaining a history of object movement. Object classification is accomplished through a multi-layer, multi-hypothesis approach that does not rely on any static features such as shape or size. Maps are stored in an efficient manner that incorporates a history of previous orientations of each object. GATMO is initialized with a static map; it subsequently changes the map over time as objects in the map change position.

Existing high-dimensional motion planning algorithms are simultaneously overpowered and underpowered. In domains sparsely populated by obstacles, the heuristics used by sampling-based planners to navigate “narrow passages” can be needlessly complex; furthermore, additional post-processing is required to remove the jerky or extraneous motions from the paths that such planners generate. In this paper, we present CHOMP, a novel method for continuous path refinement that uses covariant gradient techniques to improve the quality of sampled trajectories. Our optimization technique both optimizes higher-order dynamics and is able to converge over a wider range of input paths relative to previous path optimization strategies. In particular, we relax the collision-free feasibility prerequisite on input paths required by those strategies. As a result, CHOMP can be used as a standalone motion planner in many real-world planning queries. We demonstrate the effectiveness of our proposed method in manipulation planning for a 6-DOF robotic arm as well as in trajectory generation for a walking quadruped robot.

We present a novel approach for determining robot movements that efficiently accomplish the robot's tasks while not hindering the movements of people within the environment. Our approach models the goal-directed trajectories of pedestrians using maximum entropy inverse optimal control. The advantage of this modeling approach is the generality of its learned cost function to changes in the environment and to entirely different environments. We employ the predictions of this model of pedestrian trajectories in a novel incremental planner and quantitatively show the improvement in hindrance-sensitive robot trajectory planning provided by our approach.

We present PROCAB, an efficient method for Probabilistically Reasoning from Observed Context-Aware Behavior. It models the context-dependent utilities and underlying reasons that people take different actions. The model generalizes to unseen situations and scales to incorporate rich contextual information. We train our model using the route preferences of 25 taxi drivers demonstrated in over 100,000 miles of collected data, and demonstrate the performance of our model by inferring: (1) decision at next intersection, (2) route to known destination, and (3) destination given partially traveled route.

The widespread success of sampling-based planning algorithms stems from their ability to rapidly discover the connectivity of a configuration space. Past research has found that non-uniform sampling in the configuration space can significantly outperform uniform sampling; one important strategy is to bias the sampling distribution based on features present in the underlying workspace. In this paper, we unite several previous approaches to workspace biasing into a general framework for automatically discovering useful sampling distributions. We present a novel algorithm, based on the REINFORCE family of stochastic policy gradient algorithms, which automatically discovers a locally-optimal weighting of workspace features to produce a distribution which performs well for a given class of sampling-based motion planning queries. We present as well a novel set of workspace features that our adaptive algorithm can leverage for improved configuration space sampling. Experimental results show our algorithm to be effective across a variety of robotic platforms and high- dimensional configuration spaces.

Prior work has shown that features which appear to be biologically plausible as well as empirically useful can be found by sparse coding with a prior such as a laplacian (L1) that promotes sparsity. We show how smoother priors can pre- serve the benefits of these sparse priors while adding stability to the Maximum A-Posteriori (MAP) estimate that makes it more useful for prediction problems. Additionally, we show how to calculate the derivative of the MAP estimate effi- ciently with implicit differentiation. One prior that can be differentiated this way is KL-regularization. We demonstrate its effectiveness on a wide variety of appli- cations, and find that online optimization of the parameters of the KL-regularized model can significantly improve prediction performance.

A key challenge for autonomous navigation in cluttered outdoor environments is the reliable discrimination between obstacles that must be avoided at all costs, and lesser obstacles which the robot can drive over if necessary. Chlorophyll-rich vegetation in particular is often not an obstacle to a capable off-road vehicle, and it has long been recognized in the satellite imaging community that a simple comparison of the red and near-infrared (NIR) reflectance of a material provides a reliable technique for measuring chlorophyll content in natural scenes. This paper evaluates the effectiveness of using this chlorophyll-detection technique to improve autonomous navigation in natural, off-road environments. We demonstrate through extensive experiments that this feature has properties complementary to the color and shape descriptors traditionally used for point cloud analysis, and show significant improvement in classification performance for tasks relevant to outdoor navigation. Results are shown from field testing onboard a robot operating in off-road terrain.

Imitation learning of sequential, goal-directed behavior by standard supervised techniques is often difficult. We frame learning such behaviors as a maximum margin structured prediction problem over a space of policies. In this approach, we learn mappings from features to cost so an optimal policy in an MDP with these cost mimics the expert's behavior. Further, we demonstrate a simple, provably efficient approach to structured maximum margin learning, based on the subgradient method, that leverages existing fast algorithms for inference. Although the technique is general, it is particularly relevant in problems where A* and dynamic programming approaches make learning policies tractable in problems beyond the limitations of a QP formulation. We demonstrate our approach applied to route planning for outdoor mobile robots, where the behavior a designer wishes a planner to execute is often clear, while specifying cost functions that engender this behavior is a much more difficult task.

Long range navigation by unmanned ground vehicles continues to challenge the robotics community. Efficient navigation requires not only intelligent on-board perception and planning systems, but also the effective use of prior knowledge of the vehicle's environment. This paper describes a system for supporting unmanned ground vehicle navigation through the use of heterogeneous overhead data. Semantic information is obtained through supervised classification, and vehicle mobility is predicted from available geometric data. This approach is demonstrated and validated through over 50 kilometers of autonomous traversal through complex natural environments

The Maximum Margin Planning (MMP) (Ratliff et al., 2006) algorithm solves imitation learning problems by learning linear mappings from features to cost functions in a planning domain. The learned policy is the result of minimum-cost planning using these cost functions. These mappings are chosen so that example policies (or trajectories) given by a teacher appear to be lower cost (with a lossscaled margin) than any other policy for a given planning domain. We provide a novel approach, M M P B O O S T , based on the functional gradient descent view of boosting (Mason et al., 1999; Friedman, 1999a) that extends MMP by "boosting" in new features. This approach uses simple binary classification or regression to improve performance of MMP imitation learning, and naturally extends to the class of structured maximum margin prediction problems. (Taskar et al., 2005) Our technique is applied to navigation and planning problems for outdoor mobile robots and robotic legged locomotion.

Direct human control of multi-robot systems is limited by the cognitive ability of humans to coordinate numerous interacting components. In remote environments, such as those encountered during planetary or ocean exploration, a further limit is imposed by communication bandwidth and delay. Market based planning can give humans a higher-level interface to multi-robot systems in these scenarios. Operators provide high level tasks and attach a reward to the achievement of each task. The robots then trade these tasks through a market based mechanism. The challenge for the system designer is to create bidding algorithms for the robots that yield high overall system performance. Opportunity cost provides a nice basis for such bidding algorithms since it encapsulates all the costs and benefits we are interested in. Unfortunately, computing it can be difficult. We propose a method of learning opportunity costs in market based planners. We provide analytic results in simplified scenarios and empirical results on our FIRE simulator, which focuses on exploration of Mars by multiple, heterogeneous rovers.

We consider the scaling of the number of examples necessary to achieve good performance in distributed, cooperative, multi-agent reinforcement learning, as a function of the the number of agents n. We prove a worstcase lower bound showing that algorithms that rely solely on a global reward signal to learn policies confront a fundamental limit: They require a number of real-world examples that scales roughly linearly in the number of agents. For settings of interest with a very large number of agents, this is impractical. We demonstrate, however, that there is a class of algorithms that, by taking advantage of local reward signals in large distributed Markov Decision Processes, are able to ensure good performance with a number of samples that scales as O(log n). This makes them applicable even in settings with a very large number of agents n.

We consider the policy search approach to reinforcement learning. We show that if a “baseline distribution” is given (indicating roughly how often we expect a good policy to visit each state), then we can derive a policy search algorithm that terminates in a finite number of steps, and for which we can provide non-trivial performance guarantees. We also demonstrate this algorithm on several grid-world POMDPs, a planar biped walking robot, and a double-pole balancing problem.

Many control problems in the robotics field can be cast as partially observed Markovian decision problems (POMDPs), an optimal control formalism. Finding optimal solutions to such problems in general, however is known to be intractable. It has often been observed that in practice, simple structured controllers suffice for good sub-optimal control, and recent research in the artificial intelligence community has focused on policy search methods as techniques for finding sub-optimal controllers when such structured controllers do exist. Traditional model-based reinforcement learning algorithms make a certainty equivalence assumption on their learned models and calculate optimal policies for a maximum-likelihood Markovian model. We consider algorithms that evaluate and synthesize controllers under distributions of Markovian models. Previous work has demonstrated that algorithms that maximize mean reward with respect to model uncertainty leads to safer and more robust controllers. We consider briefly other performance criterion that emphasize robustness and exploration in the search for controllers, and note the relation with experiment design and active learning. To validate the power of the approach on a robotic application we demonstrate the presented learning control algorithm by flying an autonomous helicopter. We show that the controller learned is robust and delivers good performance in this real-world domain.