We present an option discovery algorithm that accelerates planning by minimizing the shortest distance between any two states in the MDP. The proposed algorithm produces options that approximately minimize planning time in the multi-goal setting: it is shown to be a worst case (4-alpha, 2)-approximation of the optimal option set, where alpha is the approximation ratio of the k-medians with penalties subroutine. We then present a variation, "Fast Average Options", with improved run-time and describe a general means of producing similar algorithms based on selection of a k-medians subroutine. We empirically evaluate our method on four discrete and two continuous control planning domains and show that it outperforms other leading option discovery algorithms.

Advances in reinforcement learning (RL) have led to its successful application in complex tasks with continuous state and action spaces. Despite these advances in practice, most theoretical work pertains to finite state and action spaces. We propose building a theoretical understanding of continuous state and action spaces by employing a geometric lens to understand the locally attained set of states. The set of all parametrised policies learnt through a semi-gradient based approach induce a set of attainable states in RL. We show that training dynamics of a two layer neural policy induce a low dimensional manifold of attainable states embedded in the high-dimensional nominal state space trained using an actor-critic algorithm. We prove that, under certain conditions, the dimensionality of this manifold is of the order of the dimensionality of the action space. This is the first result of its kind, linking the geometry of the state space to the dimensionality of the action space. We empirically corroborate this upper bound for four MuJoCo environments and also demonstrate the results in a toy environment with varying dimensionality. We also show the applicability of this theoretical result by introducing a local manifold learning layer to the policy and value function networks to improve the performance in control environments with very high degrees of freedom by changing one layer of the neural network to learn sparse representations.

We propose DRAGO, a novel approach for continual model-based reinforcement learning aimed at improving the incremental development of world models across a sequence of tasks that differ in their reward functions but not the state space or dynamics. DRAGO comprises two key components: Synthetic Experience Rehearsal, which leverages generative models to create synthetic experiences from past tasks, allowing the agent to reinforce previously learned dynamics without storing data, and Regaining Memories Through Exploration, which introduces an intrinsic reward mechanism to guide the agent toward revisiting relevant states from prior tasks. Together, these components enable the agent to maintain a comprehensive and continually developing world model, facilitating more effective learning and adaptation across diverse environments. Empirical evaluations demonstrate that DRAGO is able to preserve knowledge across tasks, achieving superior performance in various continual learning scenarios.

We introduce a new method that extracts knowledge from a large language model (LLM) to produce object-level plans, which describe high-level changes to object state, and uses them to bootstrap task and motion planning (TAMP). Existing work uses LLMs to directly output task plans or generate goals in representations like PDDL. However, these methods fall short because they rely on the LLM to do the actual planning or output a hard-to-satisfy goal. Our approach instead extracts knowledge from an LLM in the form of plan schemas as an object-level representation called functional object-oriented networks (FOON), from which we automatically generate PDDL subgoals. Our method markedly outperforms alternative planning strategies in completing several pick-and-place tasks in simulation.††Project Website: https://davidpaulius.github.io/olpllm/

When instructing robots, users want to flexibly express constraints, refer to arbitrary landmarks, and verify robot behavior, while robots must disambiguate instructions into specifications and ground instruction referents in the real world. To address this problem, we propose Language Instruction grounding for Motion Planning (LIMP), an approach that enables robots to verifiably follow complex, open-ended instructions in real-world environments without prebuilt semantic maps. LIMP constructs a symbolic instruction representation that reveals the robot's alignment with an instructor's intended motives and affords the synthesis of correct-by-construction robot behaviors. We conduct a large-scale evaluation of LIMP on 150 instructions across five real-world environments, demonstrating its versatility and ease of deployment in diverse, unstructured domains. LIMP performs comparably to state-of-the-art baselines on standard open-vocabulary tasks and additionally achieves a 79% success rate on complex spatiotemporal instructions, significantly outperforming baselines that only reach 38%.11See supplementary materials and demo videos at robotlimp.github.io

State uncertainty is a primary obstacle to effective long-horizon robot task planning. State uncertainty can be decomposed into spatial uncertainty—resolved using SLAM—and uncertainty about the objects in the environment, formalized as the object scouting problem and modeled using the Locally Observable Markov Decision Process (LOMDP). We introduce a new planning framework specifically designed for object scouting with LOMDPs called the Scouting Partial-Order Planner (SPOP), which exploits the characteristics of partial order and regression planning to plan around knowledge gaps the robot may have about the existence, location, and state of relevant objects in its environment. Our results highlight the benefits of partial-order planning, demonstrating its suitability for object scouting due to its ability to identify absent but task-relevant objects, and show that it outperforms comparable planners in plan length, computation time, and execution time.

We consider how to construct state abstractions compatible with a given set of abstract actions, to obtain a well-formed abstract Markov decision process (MDP). We show that the Bellman equation suggests that abstract states should represent distributions over states in the ground MDP; we characterize the conditions under which the resulting process is Markov and approximately model-preserving, derive algorithms for constructing and planning with the abstract MDP, and apply them to a visual maze task. We generalize these results to the factored actions case, characterizing the conditions that result in factored abstract states and apply the resulting algorithm to Montezuma's Revenge. These results provide a powerful and principled framework for constructing neurosymbolic abstract Markov decision processes.

We present DEPS, an end-to-end algorithm for discovering parameterized skills from expert demonstrations. Our method learns parameterized skill policies jointly with a meta-policy that selects the appropriate discrete skill and continuous parameters at each timestep. Using a combination of temporal variational inference and information-theoretic regularization methods, we address the challenge of degeneracy common in latent variable models, ensuring that the learned skills are temporally extended, semantically meaningful, and adaptable. We empirically show that learning parameterized skills from multitask expert demonstrations significantly improves generalization to unseen tasks. Our method outperforms multitask as well as skill learning baselines on both LIBERO and MetaWorld benchmarks. We also demonstrate that DEPS discovers interpretable parameterized skills, such as an object grasping skill whose continuous arguments define the grasp location.

Long-horizon decision-making tasks present significant challenges for LLM-based agents due to the need for extensive planning over multiple steps.In this paper, we propose a hierarchical framework that decomposes complex tasks into manageable subgoals, utilizing separate LLMs for subgoal prediction and lowlevel action generation.To address the challenge of creating training signals for unannotated datasets, we develop a reward model that leverages multimodal environment feedback to automatically generate reward signals.We introduce Environment Preference Optimization (EPO), a novel method that generates preference signals from the environment's feedback and uses them to train LLM-based agents.Extensive experiments on ALFRED demonstrate the state-of-the-art performance of our framework, achieving first place on the ALFRED public leaderboard and showcasing its potential to improve long-horizon decision-making in diverse environments.

We present an algorithm for skill discovery from expert demonstrations. The algorithm first utilizes Large Language Models (LLMs) to propose an initial segmentation of the trajectories. Following that, a hierarchical variational inference framework incorporates the LLM-generated segmentation information to discover reusable skills by merging trajectory segments. To further control the trade-off between compression and reusability, we introduce a novel auxiliary objective based on the Minimum Description Length principle that helps guide this skill discovery process. Our results demonstrate that agents equipped with our method are able to discover skills that help accelerate learning and outperform baseline skill learning approaches on new long-horizon tasks in BabyAI, a grid world navigation environment, as well as ALFRED, a household simulation environment.

We propose a novel model-based reinforcement learning algorithm---Dynamics Learning and predictive control with Parameterized Actions (DLPA)---for Parameterized Action Markov Decision Processes (PAMDPs). The agent learns a parameterized-action-conditioned dynamics model and plans with a modified Model Predictive Path Integral control. We theoretically quantify the difference between the generated trajectory and the optimal trajectory during planning in terms of the value they achieved through the lens of Lipschitz Continuity. Our empirical results on several standard benchmarks show that our algorithm achieves superior sample efficiency and asymptotic performance than state-of-the-art PAMDP methods.

We propose a new policy class, Composable Interaction Primitives (CIPs), specialized for learning sustained-contact manipulation skills like opening a drawer, pulling a lever, turning a wheel, or shifting gears. CIPs have two primary design goals: to minimize what must be learned by exploiting structure present in the world and the robot, and to support sequential composition by construction, so that learned skills can be used by a task-level planner. Using an ablation experiment in four simulated manipulation tasks, we show that the structure included in CIPs substantially improves the efficiency of motor skill learning. We then show that CIPs can be used for plan execution in a zero-shot fashion by sequencing learned skills. We validate our approach on real robot hardware by learning and sequencing two manipulation skills.

Deploying robots in real-world environments, such as households and manufacturing lines, requires generalization across novel task specifications without violating safety constraints. Linear temporal logic (LTL) is a widely used task specification language with a compositional grammar that naturally induces commonalities among tasks while preserving safety guarantees. However, most prior work on reinforcement learning with LTL specifications treats every new task independently, thus requiring large amounts of training data to generalize. We propose LTL-Transfer, a zero-shot transfer algorithm that composes task-agnostic skills learned during training to safely satisfy a wide variety of novel LTL task specifications. Experiments in Minecraft-inspired domains show that after training on only 50 tasks, LTL-Transfer can solve over 90% of 100 challenging unseen tasks and 100% of 300 commonly used novel tasks without violating any safety constraints. We deployed LTL-Transfer at the task-planning level of a quadruped mobile manipulator to demonstrate its zero-shot transfer ability for fetch-and-deliver and navigation tasks.

Real-world robot task planning is intractable in part due to partial observability. A common approach to reducing complexity is introducing additional structure into the decision process, such as mixed-observability, factored states, or temporally-extended actions. We propose the locally observable Markov decision process, a novel formulation that models task-level planning where uncertainty pertains to object-level attributes and where a robot has subroutines for seeking and accurately observing objects. This models sensors that are range-limited and line-of-sight—objects occluded or outside sensor range are unobserved, but the attributes of objects that fall within sensor view can be resolved via repeated observation. Our model results in a three-stage planning process: first, the robot plans using only observed objects; if that fails, it generates a target object that, if observed, could result in a feasible plan; finally, it attempts to locate and observe the target, replanning after each newly observed object. By combining LOMDPs with off-the-shelf Markov planners, we outperform state-of-the-art-solvers for both object-oriented POMDP and MDP analogues with the same task specification. We then apply the formulation to successfully solve a task on a mobile robot.

Grounding spatiotemporal navigation commands to structured task specifications enables autonomous robots to understand a broad range of natural language and solve long-horizon tasks with safety guarantees. Prior works mostly focus on grounding spatial or temporally extended language for robots. We propose Lang2LTL-2, a modular system that leverages pretrained large language and vision-language models and multimodal semantic information to ground spatiotemporal navigation commands in novel city-scaled environments without retraining. Lang2LTL-2 achieves 93.53% language grounding accuracy on a dataset of 21,780 semantically diverse natural language commands in unseen environments. We run an ablation study to validate the need for different modalities. We also show that a physical robot equipped with the same system without modification can execute 50 semantically diverse natural language commands in both indoor and outdoor environments.

Reinforcement learning algorithms typically rely on the assumption that the environment dynamics and value function can be expressed in terms of a Markovian state representation. However, when state information is only partially observable, how can an agent learn such a state representation, and how can it detect when it has found one? We introduce a metric that can accomplish both objectives, without requiring access to---or knowledge of---an underlying, unobservable state space. Our metric, the λ-discrepancy, is the difference between two distinct temporal difference (TD) value estimates, each computed using TD(λ) with a different value of λ. Since TD(λ=0) makes an implicit Markov assumption and TD(λ=1) does not, a discrepancy between these estimates is a potential indicator of a non-Markovian state representation. Indeed, we prove that the λ-discrepancy is exactly zero for all Markov decision processes and almost always non-zero for a broad class of partially observable environments. We also demonstrate empirically that, once detected, minimizing the λ-discrepancy can help with learning a memory function to mitigate the corresponding partial observability. We then train a reinforcement learning agent that simultaneously constructs two recurrent value networks with different λ parameters and minimizes the difference between them as an auxiliary loss. The approach scales to challenging partially observable domains, where the resulting agent frequently performs significantly better (and never performs worse) than a baseline recurrent agent with only a single value network.

We present Q-functionals, an alternative architecture for continuous control deep reinforcement learning. Instead of returning a single value for a state-action pair, our network transforms a state into a function that can be rapidly evaluated in parallel for many actions, allowing us to efficiently choose high-value actions through sampling. This contrasts with the typical architecture of off-policy continuous control, where a policy network is trained for the sole purpose of selecting actions from the Q-function. We represent our action-dependent Q-function as a weighted sum of basis functions (Fourier, Polynomial, etc) over the action space, where the weights are state-dependent and output by the Q-functional network. Fast sampling makes practical a variety of techniques that require Monte-Carlo integration over Q-functions, and enables action-selection strategies besides simple value-maximization. We characterize our framework, describe various implementations of Q-functionals, and demonstrate strong performance on a suite of continuous control tasks.

Principled decision-making in continuous state–action spaces is impossible without some assumptions. A common approach is to assume Lipschitz continuity of the Q-function. We show that, unfortunately, this property fails to hold in many typical domains. We propose a new coarse-grained smoothness definition that generalizes the notion of Lipschitz continuity, is more widely applicable, and allows us to compute significantly tighter bounds on Q-functions, leading to improved learning. We provide a theoretical analysis of our new smoothness definition, and discuss its implications and impact on control and exploration in continuous domains.

We introduce RLang, a domain-specific language (DSL) for communicating domain knowledge to an RL agent. Unlike existing RL DSLs that ground to $\textit{single}$ elements of a decision-making formalism (e.g., the reward function or policy), RLang can specify information about every element of a Markov decision process. We define precise syntax and grounding semantics for RLang, and provide a parser that grounds RLang programs to an algorithm-agnostic $\textit{partial}$ world model and policy that can be exploited by an RL agent. We provide a series of example RLang programs demonstrating how different RL methods can exploit the resulting knowledge, encompassing model-free and model-based tabular algorithms, policy gradient and value-based methods, hierarchical approaches, and deep methods.

We propose a novel parameterized skill-learning algorithm that aims to learn transferable parameterized skills and synthesize them into a new action space that supports efficient learning in long-horizon tasks. We propose to leverage off-policy Meta-RL combined with a trajectory-centric smoothness term to learn a set of parameterized skills. Our agent can use these learned skills to construct a three-level hierarchical framework that models a Temporally-extended Parameterized Action Markov Decision Process. We empirically demonstrate that the proposed algorithms enable an agent to solve a set of highly difficult long-horizon (obstacle-course and robot manipulation) tasks.

We propose a new method for count-based exploration in high-dimensional state spaces. Unlike previous work which relies on density models, we show that counts can be derived by averaging samples from the Rademacher distribution (or coin flips). This insight is used to set up a simple supervised learning objective which, when optimized, yields a state's visitation count. We show that our method is significantly more effective at deducing ground-truth visitation counts than previous work; when used as an exploration bonus for a model-free reinforcement learning algorithm, it outperforms existing approaches on most of 9 challenging exploration tasks, including the Atari game Montezuma's Revenge.

Conversational assistive robots can aid people, especially those with cognitive impairments, to accomplish various tasks such as cooking meals, performing exercises, or operating machines. However, to interact with people effectively, robots must recognize human plans and goals from noisy observations of human actions, even when the user acts sub-optimally. Previous works on Plan and Goal Recognition (PGR) as planning have used hierarchical task networks (HTN) to model the actor/human. However, these techniques are insufficient as they do not have user engagement via natural modes of interaction such as language. Moreover, they have no mechanisms to let users, especially those with cognitive impairments, know of a deviation from their original plan or about any sub-optimal actions taken towards their goal. We propose a novel framework for plan and goal recognition in partially observable domains—Dialogue for Goal Recognition (D4GR) enabling a robot to rectify its belief in human progress by asking clarification questions about noisy sensor data and sub-optimal human actions. We evaluate the performance of D4GR over two simulated domains—kitchen and blocks domain. With language feedback and the world state information in a hierarchical task model, we show that D4GR framework for the highest sensor noise performs 1% better than HTN in goal accuracy in both domains. For plan accuracy, D4GR outperforms by 4% in the kitchen domain and 2% in the blocks domain in comparison to HTN. The ALWAYS-ASK oracle outperforms our policy by 3% in goal recognition and 7% in plan recognition. D4GR does so by asking 68% fewer questions than an oracle baseline. We also demonstrate a real-world robot scenario in the kitchen domain, validating the improved plan and goal recognition of D4GR in a realistic setting.

It is imperative that robots can understand natural language commands issued by humans. Such commands typically contain verbs that signify what action should be performed on a given object and that are applicable to many objects. We propose a method for generalizing manipulation skills to novel objects using verbs. Our method learns a probabilistic classifier that determines whether a given object trajectory can be described by a specific verb. We show that this classifier accurately generalizes to novel object categories with an average accuracy of 76.69% across 13 object categories and 14 verbs. We then perform policy search over the object kinematics to find an object trajectory that maximizes classifier prediction for a given verb. Our method allows a robot to generate a trajectory for a novel object based on a verb, which can then be used as input to a motion planner. We show that our model can generate trajectories that are usable for executing five verb commands applied to novel instances of two different object categories on a real robot.

Dynamic movement primitives are widely used for learning skills that can be demonstrated to a robot by a skilled human or controller. While their generalization capabilities and simple formulation make them very appealing to use, they possess no strong guarantees to satisfy operational safety constraints for a task. We present constrained dynamic movement primitives (CDMPs), which can allow for positional constraint satisfaction in the robot workspace. Our method solves a non-linear optimization to perturb an existing DMP's forcing weights to admit a Zeroing Barrier Function (ZBF), which certifies positional workspace constraint satisfaction. We demonstrate our approach under different positional constraints on the end-effector movement on multiple physical robots, such as obstacle avoidance and workspace limitations.

An agent learning an option in hierarchical reinforcement learning must solve three problems: identify the option's subgoal (termination condition), learn a policy, and learn where that policy will succeed (initiation set). The termination condition is typically identified first, but the option policy and initiation set must be learned simultaneously, which is challenging because the initiation set depends on the option policy, which changes as the agent learns. Consequently, data obtained from option execution becomes invalid over time, leading to an inaccurate initiation set that subsequently harms downstream task performance. We highlight three issues---data non-stationarity, temporal credit assignment, and pessimism---specific to learning initiation sets, and propose to address them using tools from off-policy value estimation and classification. We show that our method learns higher-quality initiation sets faster than existing methods (in MiniGrid and Montezuma's Revenge), can automatically discover promising grasps for robot manipulation (in Robosuite), and improves the performance of a state-of-the-art option discovery method in a challenging maze navigation task in MuJoCo.

Optimistic initialization underpins many theoretically sound exploration schemes in tabular domains; however, in the deep function approximation setting, optimism can quickly disappear if initialized naively. We propose a framework for more effectively incorporating optimistic initialization into reinforcement learning for continuous control. Our approach uses metric information about the state-action space to estimate which transitions are still unexplored, and explicitly maintains the initial Q-value optimism for the corresponding state-action pairs. We also develop methods for efficiently approximating these training objectives, and for incorporating domain knowledge into the optimistic envelope to improve sample efficiency. We empirically evaluate these approaches on a variety of hard exploration problems in continuous control, where our method outperforms existing exploration techniques.

Manipulating an articulated object requires perceiving its kinematic hierarchy: its parts, how each can move, and how those motions are coupled. Previous work has explored perception for kinematics, but none infers a complete kinematic hierarchy on never-before-seen object instances, without relying on a schema or template. We present a novel perception system that achieves this goal. Our system infers the moving parts of an object and the kinematic couplings that relate them. To infer parts, it uses a point cloud instance segmentation neural network and to infer kinematic hierarchies, it uses a graph neural network to predict the existence, direction, and type of edges (i.e. joints) that relate the inferred parts. We train these networks using simulated scans of synthetic 3D models. We evaluate our system on simulated scans of 3D objects, and we demonstrate a proof-of-concept use of our system to drive real-world robotic manipulation.

Robots that process navigation instructions in large outdoor environments will need to operate at different levels of abstraction. For example, a land-surveying aerial robot receiving the instruction “go to Boston and go through the state forest on the way” must reason about a long-range goal like “go to Boston” while also processing a finer-grained constraint like “go through the state forest.” Existing approaches struggle to plan such commands because of the immense number of locations and constraints that can be expressed in language. We introduce a hierarchical representation of outdoor environments and a planning approach that dynamically compacts the robot's state space to enable tractable planning in city and state-scale environments. Our approach leverages natural abstractions in real-world map data, coupled with abstractions generated from users' instructions, to generate filtered environment views that accelerate planning while supporting a robot's ability to obey complex temporal goals and constraints at different levels of abstraction. We evaluate our approach on seven templates of LTLJ formulas and in an 80 kilometer-radius environment containing over 250,000 locations downloaded from OpenStreetMap. The results show our approach enables planning in seconds or minutes in a large outdoor environment while still satisfying the task specification.

Recent work on using natural language to specify commands to robots has grounded that language to LTL. However, mapping natural language task specifications to LTL task specifications using language models require probability distributions over finite vocabulary. Existing state-of-the-art methods have extended this finite vocabulary to include unseen terms from the input sequence to improve output generalization. However, novel out-of-vocabulary atomic propositions cannot be generated using these methods. To overcome this, we introduce an intermediate contextual query representation which can be learned from single positive task specification examples, associating a contextual query with an LTL template. We demonstrate that this intermediate representation allows for generalization over unseen object references, assuming accurate groundings are available. We compare our method of mapping natural language task specifications to intermediate contextual queries against state-of-the-art CopyNet models capable of translating natural language to LTL, by evaluating whether correct LTL for manipulation and navigation task specifications can be output, and show that our method outperforms the CopyNet model on unseen object references. We demonstrate that the grounded LTL our method outputs can be used for planning in a simulated OO-MDP environment. Finally, we discuss some common failure modes encountered when translating natural language task specifications to grounded LTL.

Variable impedance control in operation-space is a promising approach to learning contact-rich manipulation behaviors. One of the main challenges with this approach is producing a manipulation behavior that ensures the safety of the arm and the environment. Such behavior is typically implemented via a reward function that penalizes unsafe actions (e.g. obstacle collision, joint limit extension), but that approach is not always effective and does not result in behaviors that can be reused in slightly different environments. We show how to combine Riemannian Motion Policies, a class of policies that dynamically generate motion in the presence of safety and collision constraints, with variable impedance operation-space control to learn safer contact-rich manipulation behaviors.

In realistic applications of object search, robots will need to locate target objects in complex environments while coping with unreliable sensors, especially for small or hard-to-detect objects. In such settings, correlational information can be valuable for planning efficiently. Previous approaches that consider correlational information typically resort to ad-hoc, greedy search strategies. We introduce the Correlational Object Search POMDP (COS-POMDP), which models correlations while preserving optimal solutions with a reduced state space. We propose a hierarchical planning algorithm to scale up COS-POMDPs for practical domains. Our evaluation, conducted with the AI2-THOR household simulator and the YOLOv5 object detector, shows that our method finds objects more successfully and efficiently compared to baselines, particularly for hard-to-detect objects such as srub brush and remote control.

We propose a model-based lifelong reinforcement-learning approach that estimates a hierarchical Bayesian posterior distilling the common structure shared across different tasks. The learned posterior combined with a sample-based Bayesian exploration procedure increases the sample efficiency of learning across a family of related tasks. We first derive an analysis of the relationship between the sample complexity and the initialization quality of the posterior in the finite MDP setting. We next scale the approach to continuous-state domains by introducing a Variational Bayesian Lifelong Reinforcement Learning algorithm that can be combined with recent model-based deep RL methods, and that exhibits backward transfer. Experimental results on several challenging domains show that our algorithms achieve both better forward and backward transfer performance than state-of-the-art lifelong RL methods.

AlphaZero, an approach to reinforcement learning that couples neural networks and Monte Carlo tree search (MCTS), has produced state-of-the-art strategies for traditional board games like chess, Go, shogi, and Hex. While researchers and game commentators have suggested that AlphaZero uses concepts that humans consider important, it is unclear how these concepts are captured in the network. We investigate AlphaZero's internal representations in the game of Hex using two evaluation techniques from natural language processing (NLP): model probing and behavioral tests. In doing so, we introduce several new evaluation tools to the RL community, and illustrate how evaluations other than task performance can be used to provide a more complete picture of a model's strengths and weaknesses. Our analyses in the game of Hex reveal interesting patterns and generate some testable hypotheses about how such models learn in general. For example, we find that the MCTS discovers concepts before the neural network learns to encode them. We also find that concepts related to short-term end-game planning are best encoded in the final layers of the model, whereas concepts related to long-term planning are encoded in the middle layers of the model.

Deep neural networks can approximate functions on different types of data, from images to graphs, with varied underlying structure. This underlying structure can be viewed as the geometry of the data manifold. By extending recent advances in the theoretical understanding of neural networks, we study how a randomly initialized neural network with piecewise linear activation splits the data manifold into regions where the neural network behaves as a linear function. We derive bounds on the density of boundary of linear regions and the distance to these boundaries on the data manifold. This leads to insights into the expressivity of randomly initialized deep neural networks on non-Euclidean data sets. We empirically corroborate our theoretical results using a toy supervised learning problem. Our experiments demonstrate that number of linear regions varies across manifolds and the results hold with changing neural network architectures. We further demonstrate how the complexity of linear regions is different on the low dimensional manifold of images as compared to the Euclidean space, using the MetFaces dataset.

A core operation in reinforcement learning (RL) is finding an action that is optimal with respect to a learned value function. This operation is often challenging when the learned value function takes continuous actions as input. We introduce deep radial-basis value functions (RBVFs): value functions learned using a deep network with a radial-basis function (RBF) output layer. We show that the maximum action-value with respect to a deep RBVF can be approximated easily and accurately. Moreover, deep RBVFs can represent any true value function owing to their support for universal function approximation. We extend the standard DQN algorithm to continuous control by endowing the agent with a deep RBVF. We show that the resultant agent, called RBF-DQN, significantly outperforms value-function-only baselines, and is competitive with state-of-the-art actor-critic algorithms.

We introduce Wasserstein Adversarial Proximal Policy Optimization (WAPPO), a novel algorithm for visual transfer in Reinforcement Learning that explicitly learns to align the distributions of extracted features between a source and target task. WAPPO approximates and minimizes the Wasserstein-1 distance between the distributions of features from source and target domains via a novel Wasserstein Confusion objective. WAPPO outperforms the prior state-of-the-art in visual transfer and successfully transfers policies across Visual Cartpole and both the easy and hard settings of of 16 OpenAI Procgen environments.

We introduce a new skill-discovery algorithm that builds a discrete graph representation of large continuous MDPs, where nodes correspond to skill subgoals and the edges to skill policies. The agent constructs this graph during an unsupervised training phase where it interleaves discovering skills and planning using them to gain coverage over ever-increasing portions of the state-space. Given a novel goal at test time, the agent plans with the acquired skill graph to reach a nearby state, then switches to learning to reach the goal. We show that the resulting algorithm, Deep Skill Graphs, outperforms both flat and existing hierarchical reinforcement learning methods on four difficult continuous control tasks.

Robots must reason about pushing and grasping in order to engage in flexible manipulation in cluttered environments. Earlier works on learning pushing and grasping only consider each operation in isolation or are limited to top-down grasping and bin-picking. We train a robot to learn joint planar pushing and 6-degree-of-freedom (6-DoF) grasping policies by self-supervision. Two separate deep neural networks are trained to map from 3D visual observations to actions with a Q-learning framework. With collaborative pushes and expanded grasping action space, our system can deal with cluttered scenes with a wide variety of objects (e.g. grasping a plate from the side after pushing away surrounding obstacles). We compare our system to the state-of-the-art baseline model VPG [1] in simulation and outperform it with 10% higher action efficiency and 20% higher grasp success rate. We then demonstrate our system on a KUKA LBR iiwa arm with a Robotiq 3-finger gripper.

The difficulty of deterministic planning increases exponentially with search-tree depth. Black-box planning presents an even greater challenge, since planners must operate without an explicit model of the domain. Heuristics can make search more efficient, but goal-aware heuristics for black-box planning usually rely on goal counting, which is often quite uninformative. In this work, we show how to overcome this limitation by discovering macro-actions that make the goal-count heuristic more accurate. Our approach searches for macro-actions with focused effects (i.e. macros that modify only a small number of state variables), which align well with the assumptions made by the goal-count heuristic. Focused macros dramatically improve black-box planning efficiency across a wide range of planning domains, sometimes beating even state-of-the-art planners with access to a full domain model.

Hierarchical reinforcement learning (HRL) is only effective for long-horizon problems when high-level skills can be reliably sequentially executed. Unfortunately, learning reliably composable skills is difficult, because all the components of every skill are constantly changing during learning. We propose three methods for improving the composability of learned skills: representing skill initiation regions using a combination of pessimistic and optimistic classifiers; learning re-targetable policies that are robust to non-stationary subgoal regions; and learning robust option policies using model-based RL. We test these improvements on four sparse-reward maze navigation tasks involving a simulated quadrupedal robot. Each method successively improves the robustness of a baseline skill discovery method, substantially outperforming state-of-the-art flat and hierarchical methods.

Learning a robot motor skill from scratch is impractically slow; so much so that in practice, learning must typically be bootstrapped using human demonstration. However, relying on human demonstration necessarily degrades the autonomy of robots that must learn a wide variety of skills over their operational lifetimes. We propose using kinematic motion planning as a completely autonomous, sample efficient way to bootstrap motor skill learning for object manipulation. We demonstrate the use of motion planners to bootstrap motor skills in two complex object manipulation scenarios with different policy representations: opening a drawer with a dynamic movement primitive representation, and closing a microwave door with a deep neural network policy. We also show how our method can bootstrap a motor skill for the challenging dynamic task of learning to hit a ball off a tee, where a kinematic plan based on treating the scene as static is insufficient to solve the task, but sufficient to bootstrap a more dynamic policy. In all three cases, our method is competitive with human-demonstrated initialization, and significantly out-performs starting with a random policy. This approach enables robots to to efficiently and autonomously learn motor policies for dynamic tasks without human demonstration.

We propose an approach to multi-modal grasp detection that jointly predicts the probabilities that several types of grasps succeed at a given grasp pose. Given a partial point cloud of a scene, the algorithm proposes a set of feasible grasp candidates, then estimates the probabilities that a grasp of each type would succeed at each candidate pose. Predicting grasp success probabilities directly from point clouds makes our approach agnostic to the number and placement of depth sensors at execution time. We evaluate our system both in simulation and on a real robot with a Robotiq 3-Finger Adaptive Gripper and compare our network against several baselines that perform fewer types of grasps. Our experiments show that a system that explicitly models grasp type achieves an object retrieval rate 8.5% higher in a complex cluttered environment than our highest-performing baseline.

Robots operating in households must find objects on shelves, under tables, and in cupboards. In such environments, it is crucial to search efficiently at 3D scale while coping with limited field of view and the complexity of searching for multiple objects. Principled approaches to object search frequently use Partially Observable Markov Decision Process (POMDP) as the underlying framework for computing search strategies, but constrain the search space in 2D. In this paper, we present a POMDP formulation for multi-object search in a 3D region with a frustum-shaped field-of-view. To efficiently solve this POMDP, we propose a multi-resolution planning algorithm based on online Monte-Carlo tree search. In this approach, we design a novel octree-based belief representation to capture uncertainty of the target objects at different resolution levels, then derive abstract POMDPs at lower resolutions with dramatically smaller state and observation spaces. Evaluation in a simulated 3D domain shows that our approach finds objects more efficiently and successfully compared to a set of baselines without resolution hierarchy in larger instances under the same computational requirement. We demonstrate our approach on a mobile robot to find objects placed at different heights in two 10m2×2m regions by moving its base and actuating its torso.

A fundamental assumption of reinforcement learning in Markov decision processes (MDPs) is that the relevant decision process is, in fact, Markov. However, when MDPs have rich observations, agents typically learn by way of an abstract state representation, and such representations are not guaranteed to preserve the Markov property. We introduce a novel set of conditions and prove that they are sufficient for learning a Markov abstract state representation. We then describe a practical training procedure that combines inverse model estimation and temporal contrastive learning to learn an abstraction that approximately satisfies these conditions. Our novel training objective is compatible with both online and offline training: it does not require a reward signal, but agents can capitalize on reward information when available. We empirically evaluate our approach on a visual gridworld domain and a set of continuous control benchmarks. Our approach learns representations that capture the underlying structure of the domain and lead to improved sample efficiency over state-of-the-art deep reinforcement learning with visual features---often matching or exceeding the performance achieved with hand-designed compact state information.

A key challenge in intelligent robotics is creating robots that are capable of directly interacting with the world around them to achieve their goals. The last decade has seen substantial growth in research on the problem of robot manipulation, which aims to exploit the increasing availability of affordable robot arms and grippers to create robots capable of directly interacting with the world to achieve their goals. Learning will be central to such autonomous systems, as the real world contains too much variation for a robot to expect to have an accurate model of its environment, the objects in it, or the skills required to manipulate them, in advance. We aim to survey a representative subset of that research which uses machine learning for manipulation. We describe a formalization of the robot manipulation learning problem that synthesizes existing research into a single coherent framework and highlight the many remaining research opportunities and challenges.

We propose an abstraction method for open-world environments expressed as Factored Markov Decision Processes (FMDPs) with very large state and action spaces. Our method prunes state and action variables that are irrelevant to the optimal value function on the state subspace the agent would visit when following any optimal policy from the initial state. This method thus enables tractable fast planning within large open-world FMDPs.

We present a framework for autonomously learning a portable representation that describes a collection of low-level continuous environments. We show that these abstract representations can be learned in a task-independent egocentric space specific to the agent that, when grounded with problem-specific information, are provably sufficient for planning. We demonstrate transfer in two different domains, where an agent learns a portable, task-independent symbolic vocabulary, as well as operators expressed in that vocabulary, and then learns to instantiate those operators on a per-task basis. This reduces the number of samples required to learn a representation of a new task.

Precomputed roadmaps can enable effective multi-query motion planning: a roadmap can be built for a robot as if no obstacles were present, and then after edges invalidated by obstacles observed at query time are deleted, path search through the remaining roadmap returns a collision-free plan. However, large roadmaps are memory intensive to store, and can be too slow for practical use. We present an algorithm for compressing a large roadmap so that the collision detection phase fits into a computational budget, while retaining a high probability of finding high-quality paths. Our algorithm adapts work from graph theory and data mining by treating roadmaps as unreliable networks, where the probability of edge failure models the probability of a query-time obstacle causing a collision. We experimentally evaluate the quality of the resulting roadmaps in a suite of four motion planning benchmarks.

We propose Action-Oriented Semantic Maps (AOSMs), a representation that enables a robot to acquire object manipulation behaviors and semantic information about the environment from a human teacher with a Mixed Reality Head-Mounted Display (MR-HMD). AOSMs are a representation that captures both: a) high-level object manipulation actions in an object class's local frame, and b) semantic representations of objects in the robot's global map that are grounded for navigation. Humans can use a MR-HMD to teach the agent the information necessary for planning object manipulation and navigation actions by interacting with virtual 3D meshes overlaid on the physical workspace. We demonstrate that our system enables users to quickly and accurately teach a robot the knowledge required to autonomously plan and execute three household tasks: picking up a bottle and throwing it in the trash, closing a sink faucet, and flipping a light switch off.

Ophthalmic microsurgery is technically difficult because the scale of required surgical tool manipulations challenge the limits of the surgeon's visual acuity, sensory perception, and physical dexterity. Intraoperative optical coherence tomography (OCT) imaging with micrometer-scale resolution is increasingly being used to monitor and provide enhanced real-time visualization of ophthalmic surgical maneuvers, but surgeons still face physical limitations when manipulating instruments inside the eye. Autonomously controlled robots are one avenue for overcoming these physical limitations. We demonstrate the feasibility of using learning from demonstration and reinforcement learning with an industrial robot to perform OCT-guided corneal needle insertions in an ex vivo model of deep anterior lamellar keratoplasty (DALK) surgery. Our reinforcement learning agent trained on ex vivo human corneas, then outperformed surgical fellows in reaching a target needle insertion depth in mock corneal surgery trials. This work shows the combination of learning from demonstration and reinforcement learning is a viable option for performing OCT guided robotic ophthalmic surgery.

We have designed a programmable architecture to accelerate collision detection and graph search, two of the principal components of robotic motion planning. The programmability enables the architecture to be applied to a wide range of different robots and motion planning applications. We present the architecture of our accelerator and describe and evaluate its microarchitecture implementation.

We formalize the problem of selecting the optimal set of options for planning as that of computing the smallest set of options so that planning converges in less than a given maximum of value-iteration passes. We first show that the problem is $\NP$-hard, even if the task is constrained to be deterministic—the first such complexity result for option discovery. We then present the first polynomial-time boundedly suboptimal approximation algorithm for this setting, and empirically evaluate it against both the optimal options and a representative collection of heuristic approaches in simple grid-based domains.

One of the main challenges in reinforcement learning is solving tasks with sparse reward. We show that the difficulty of discovering a distant rewarding state in an MDP is bounded by the expected cover time of a random walk over the graph induced by the MDP’s transition dynamics. We therefore propose to accelerate exploration by constructing options that minimize cover time. We introduce a new option discovery algorithm that diminishes the expected cover time by connecting the most distant states in the state-space graph with options. We show empirically that the proposed algorithm improves learning in several domains with sparse rewards.

Security is particularly important in robotics, as robots can directly perceive and affect the physical world. We describe the results of a scan of the entire IPv4 address space of the Internet for instances of the Robot Operating System (ROS), a widely used robotics software platform. We identified a number of hosts supporting ROS that are exposed to the public Internet, thereby allowing anyone to access robotic sensors and actuators. As a proof of concept, and with the consent of the relevant researchers, we were able to read image sensor information from and actuate a physical robot present in a research lab in an American university. This paper gives an overview of our findings, including our methodology, the geographic distribution of publicly-accessible platforms, the sorts of sensor and actuator data that is available, and the different kinds of robots and sensors that our scan uncovered. Additionally, we offer recommendations on best practices to mitigate these security issues in the future.

Mixed Reality (MR) is a promising interface for robot programming because it can project an immersive 3D visualization of a robot's intended movement onto the real world. MR can also support hand gestures, which provide an intuitive way for users to construct and modify robot motions. We present a Mixed Reality Head-Mounted Display (MRHMD) interface that enables end-users to easily create and edit robot motions using waypoints. We describe a user study where 20 participants were asked to program a robot arm using 2D and MR interfaces to perform two pick-and-place tasks. In the primitive task, participants created typical pickand-place programs. In the adapted task, participants adapted their primitive programs to address a more complex pickand-place scenario, which included obstacles and conditional reasoning. Compared to the 2D interface, a higher number of users were able to complete both tasks in significantly less time, and reported experiencing lower cognitive workload, higher usability, and higher naturalness with the MR-HMD interface.

Deep Q-Network (DQN) is an algorithm that achieves human-level performance in complex domains like Atari games. One of the important elements of DQN is its use of a target network, which is necessary to stabilize learning. We argue that using a target network is incompatible with online reinforcement learning, and it is possible to achieve faster and more stable learning without a target network when we use Mellowmax, an alternative softmax operator. We derive novel properties of Mellowmax, and empirically show that the combination of DQN and Mellowmax, but without a target network, outperforms DQN with a target network.

We present a feedback motion planning algorithm, Bounded-Error LQR-Trees, that leverages reinforcement learning theory to find a policy with a bounded amount of error. The algorithm composes locally valid linear-quadratic regulators (LQR) into a nonlinear controller, similar to how LQR-Trees constructs its policy, but minimizes the cost of the constructed policy by minimizing the Bellman Residual, which is estimated in the overlapping regions of LQR controllers. We prove a sample-based upper bound on the true Bellman Residual, and demonstrate a five-fold reduction in cost over previous methods on a simple underactuated nonlinear system.

We develop a system to disambiguate object instances within the same class based on simple physical descriptions. The system takes as input a natural language phrase and a depth image containing a segmented object and predicts how similar the observed object is to the object described by the phrase. Our system is designed to learn from only a small amount of human-labeled language data and generalize to viewpoints not represented in the language-annotated depth image training set. By decoupling 3D shape representation from language representation, this method is able to ground language to novel objects using a small amount of language-annotated depth-data and a larger corpus of unlabeled 3D object meshes, even when these objects are partially observed from unusual viewpoints. Our system is able to disambiguate between novel objects, observed via depth images, based on natural language descriptions. Our method also enables viewpoint transfer; trained on human-annotated data on a small set of depth images captured from frontal viewpoints, our system successfully predicted object attributes from rear views despite having no such depth images in its training set. Finally, we demonstrate our approach on a Baxter robot, enabling it to pick specific objects based on human-provided natural language descriptions.

Decentralized partially observable Markov decision processes (Dec-POMDPs) are general models for decentralized multi-agent decision making under uncertainty. However, they typically model a problem at a low level of granularity, where each agent's actions are primitive operations lasting exactly one time step. We address the case where each agent has macro-actions: temporally extended actions that may require different amounts of time to execute. We model macro-actions as options in a Dec-POMDP, focusing on actions that depend only on information directly available to the agent during execution. Therefore, we model systems where coordination decisions only occur at the level of deciding which macro-actions to execute. The core technical difficulty in this setting is that the options chosen by each agent no longer terminate at the same time. We extend three leading Dec-POMDP algorithms for policy generation to the macro-action case, and demonstrate their effectiveness in both standard benchmarks and a multi-robot coordination problem. The results show that our new algorithms retain agent coordination while allowing high-quality solutions to be generated for significantly longer horizons and larger state-spaces than previous Dec-POMDP methods. Furthermore, in the multi-robot domain, we show that, in contrast to most existing methods that are specialized to a particular problem class, our approach can synthesize control policies that exploit opportunities for coordination while balancing uncertainty, sensor information, and information about other agents.

We consider the problem of how best to use prior experience to bootstrap lifelong learning, where an agent faces a series of task instances drawn from some task distribution. First, we identify the initial policy that optimizes expected performance over the distribution of tasks for increasingly complex classes of policy and task distributions. We empirically demonstrate the relative performance of each policy class’ optimal element in a variety of simple task distributions. We then consider value-function initialization methods that preserve PAC guarantees while simultaneously minimizing the learning required in two learning algorithms, yielding MaxQInit, a practical new method for value-function-based transfer. We show that MaxQInit performs well in simple lifelong RL experiments.

A critical capability required for generally intelligent robot behavior is the ability to sequence motor skills to reach a goal. This requires a (typically abstract) representation that supports goal-directed planning, which raises the question of how to construct such a representation. Previous work has addressed this question in the context of simple black-box motor skills, which are insufficiently flexible to support the wide range of behavior required of intelligent robots. We now extend that work to include parametrized motor skills, where a robot must both select an action to execute and also decide how to parametrize it. We show how to construct a representation suitable for planning with parametrized motor skills, and specify conditions which are sufficient to separate the selection of motor skills from the parametrization of those skills. Our method results in a simple discrete abstract representation for planning followed by a parameter selection process that operates on a fixed plan. We first demonstrate learning this representation in a virtual domain based on Angry Birds and then learn an abstract symbolic representation for a robot manipulation task.

We introduce Hybrid Bayesian Eigenobjects (HBEOs), a novel representation for 3D objects designed to allow a robot to jointly estimate the pose, class, and full 3D geometry of a novel object observed from a single viewpoint in a single practical framework. By combining both linear subspace methods and deep convolutional prediction, HBEOs efficiently learn nonlinear object representations without directly regressing into high-dimensional space. HBEOs also remove the onerous and generally impractical necessity of input data voxelization prior to inference. We experimentally evaluate the suitability of HBEOs to the challenging task of joint pose, class, and shape inference on novel objects and show that, compared to preceding work, HBEOs offer dramatically improved performance in all three tasks along with several orders of magnitude faster runtime performance.

We consider the problem of constructing abstract representations for planning in high-dimensional, continuous environments. We assume an agent equipped with a collection of high-level actions, and construct representations provably capable of evaluating plans composed of sequences of those actions. We first consider the deterministic planning case, and show that the relevant computation involves set operations performed over sets of states. We define the specific collection of sets that is necessary and sufficient for planning, and use them to construct a grounded abstract symbolic representation that is provably suitable for deterministic planning. The resulting representation can be expressed in PDDL, a canonical high-level planning domain language; we construct such a representation for the Playroom domain and solve it in milliseconds using an off-the-shelf planner. We then consider probabilistic planning, which we show requires generalizing from sets of states to distributions over states. We identify the specific distributions required for planning, and use them to construct a grounded abstract symbolic representation that correctly estimates the expected reward and probability of success of any plan. In addition, we show that learning the relevant probability distributions corresponds to specific instances of probabilistic density estimation and probabilistic classification. We construct an agent that autonomously learns the correct abstract representation of a computer game domain, and rapidly solves it. Finally, we apply these techniques to create a physical robot system that autonomously learns its own symbolic representation of a mobile manipulation task directly from sensorimotor data---point clouds, map locations, and joint angles---and then plans using that representation. Together, these results establish a principled link between high-level actions and abstract representations, a concrete theoretical foundation for constructing abstract representations with provable properties, and a practical mechanism for autonomously learning abstract high-level representations.

Monte Carlo Tree Search (MCTS) is a family of directed search algorithms that has gained widespread attention in recent years. Despite the vast amount of research into MCTS, the effect of modifications on the algorithm, as well as the manner in which it performs in various domains, is still not yet fully known. In particular, the effect of using knowledge-heavy rollouts in MCTS still remains poorly understood, with surprising results demonstrating that better-informed rollouts often result in worse-performing agents. We present experimental evidence suggesting that, under certain smoothness conditions, uniformly random simulation policies preserve the ordering over action preferences. This explains the success of MCTS despite its common use of these rollouts to evaluate states. We further analyse non-uniformly random rollout policies and describe conditions under which they offer improved performance.

An intriguing application of transfer learning emerges when tasks arise with similar, but not identical, dynamics. Hidden Parameter Markov Decision Processes (HiP-MDP) embed these tasks into a low-dimensional space; given the embedding parameters one can identify the MDP for a particular task. However, the original formulation of HiP-MDP had a critical flaw: the embedding uncertainty was modeled independently of the agent's state uncertainty, requiring an arduous training procedure. In this work, we apply a Gaussian Process latent variable model to jointly model the dynamics and the embedding, leading to a more elegant formulation, one that allows for better uncertainty quantification and thus more robust transfer.

We introduce an online active exploration algorithm for data-efficiently learning an abstract symbolic model of an environment. Our algorithm is divided into two parts: the first part quickly generates an intermediate Bayesian symbolic model from the data that the agent has collected so far, which the agent can then use along with the second part to guide its future exploration towards regions of the state space that the model is uncertain about. We show that our algorithm outperforms random and greedy exploration policies on two different computer game domains. The first domain is an Asteroids-inspired game with complex dynamics but basic logical structure. The second is the Treasure Game, with simpler dynamics but more complex logical structure.

We introduce a new formulation of the Hidden Parameter Markov Decision Process (HiP-MDP), a framework for modeling families of related tasks using low-dimensional latent embeddings. Our new framework correctly models the joint uncertainty in the latent parameters and the state space. We also replace the original Gaussian Process-based model with a Bayesian Neural Network, enabling more scalable inference. Thus, we expand the scope of the HiP-MDP to applications with higher dimensions and more complex dynamics.

We introduce a model-free algorithm for learning in Markov decision processes with parameterized actions—discrete actions with continuous parameters. At each step the agent must select both which action to use and which parameters to use with that action. We introduce the Q-PAMDP algorithm for learning in these domains, show that it converges to a local optimum, and compare it to direct policy search in the goal-scoring and Platform domains.

We have developed a hardware accelerator for motion planning, a critical operation in robotics. In this paper, we present the microarchitecture of our accelerator and describe a prototype implementation on an FPGA. We experimentally show that the accelerator improves performance by three orders of magnitude and improves power consumption by more than one order of magnitude. These gains are achieved through careful hardware/software co-design. We modify conventional motion planning algorithms to aggressively precompute collision data, as well as implement a microarchitecture that leverages the parallelism present in the problem.

This paper presents a probabilistic framework for synthesizing control policies for general multi-robot systems that is based on decentralized partially observable Markov decision processes (Dec-POMDPs). Dec-POMDPs are a general model of decision-making where a team of agents must cooperate to optimize a shared objective in the presence of uncertainty. Dec-POMDPs also consider communication limitations, so execution is decentralized. While Dec-POMDPs are typically intractable to solve for real-world problems, recent research on the use of macro-actions in Dec-POMDPs has significantly increased the size of problem that can be practically solved. We show that, in contrast to most existing methods that are specialized to a particular problem class, our approach can synthesize control policies that exploit any opportunities for coordination that are present in the problem, while balancing uncertainty, sensor information, and information about other agents. We use three variants of a warehouse task to show that a single planner of this type can generate cooperative behavior using task allocation, direct communication, and signaling, as appropriate. This demonstrates that our algorithmic framework can automatically optimize control and communication policies for complex multi-robot systems.

We present a method for segmenting a set of unstructured demonstration trajectories to discover reusable skills using inverse reinforcement learning (IRL). Each skill is characterised by a latent reward function which the demonstrator is assumed to be optimizing. The skill boundaries and the number of skills making up each demonstration are unknown. We use a Bayesian nonparametric approach to propose skill segmentations and maximum entropy inverse reinforcement learning to infer reward functions from the segments. This method produces a set of Markov Decision Processes (MDPs) that best describe the input trajectories. We evaluate this approach in a car driving domain and a simulated quadcopter obstacle course, showing that it is able to recover demonstrated skills more effectively than existing methods.

We propose the Ω-return as an alternative to the λ-return currently used by the TD(λ) family of algorithms. The benefit of the Ω-return is that it accounts for the correlation of different length returns. Because it is difficult to compute exactly, we suggest one way of approximating the Ω-return. We provide empirical studies that suggest that it is superior to the λ-return and γ-return for a variety of problems.

We applied a policy search algorithm to the problem of optimizing a start-stop controller—a controller used in a car to turn off the vehicle’s engine, and thus save energy, when the vehicle comes to a temporary halt. We were able to improve the existing policy by approximately 12% using real driver trace data. We also experimented with using multiple policies, and found that doing so could lead to a further 8% improvement if we could determine which policy to apply at each stop. The driver’s behaviors before stopping were found to be uncorrelated with the policy that performed best; however, further experimentation showed that the driver’s behavior during the stop may be more useful, suggesting a useful direction for adding complexity to the underlying start-stop policy.

We consider the problem of constructing a symbolic description of a continuous, low-level environment for use in planning. We show that symbols that can represent the preconditions and effects of an agent's actions are both necessary and sufficient for high-level planning. This eliminates the symbol design problem when a representation must be constructed in advance, and in principle enables an agent to autonomously learn its own symbolic representations. The resulting representation can be converted into PDDL, a canonical high-level planning representation that enables very fast planning.

We introduce a method for actively learning parameterized skills. Parameterized skills are flexible behaviors that can solve any task drawn from a distribution of parameterized reinforcement learning problems. Approaches to learning such skills have been proposed, but limited attention has been given to identifying which training tasks allow for rapid skill acquisition. We construct a non-parametric Bayesian model of skill performance and derive analytical expressions for a novel acquisition criterion capable of identifying tasks that maximize expected improvement in skill performance. We also introduce a spatiotemporal kernel tailored for non-stationary skill performance models. The proposed method is agnostic to policy and skill representation and scales independently of task dimensionality. We evaluate it on a non-linear simulated catapult control problem over arbitrarily mountainous terrains.

We demonstrate a sample-efficient method for constructing reusable parameterized skills that can solve families of related motor tasks. Our method uses learned policies to analyze the policy space topology and learn a set of regression models which, given a novel task, appropriately parameterizes an underlying low-level controller. By identifying the disjoint charts that compose the policy manifold, the method can separately model the qualitatively different sub-skills required for solving distinct classes of tasks. Such sub-skills are useful because they can be treated as new discrete, specialized actions by higher-level planning processes. We also propose a method for reusing seemingly unsuccessful policies as additional, valid training samples for synthesizing the skill, thus accelerating learning. We evaluate our method on a humanoid iCub robot tasked with learning to accurately throw plastic balls at parameterized target locations.

We propose a new method for applying RRT* to kinodynamic motion planning problems by using finite-horizon linear quadratic regulation (LQR) to measure cost and to extend the tree. First, we introduce the method in the context of arbitrary affine dynamical systems with quadratic costs. For these systems, the algorithm is shown to converge to optimal solutions almost surely. Second, we extend the algorithm to non-linear systems with non-quadratic costs, and demonstrate its performance experimentally.

The RRT* algorithm has recently been proposed as an optimal extension to the standard RRT algorithm [1]. However, like RRT, RRT* is difficult to apply in problems with complicated or underactuated dynamics because it requires the design of a two domain-specific extension heuristics: a distance metric and node extension method. We propose automatically deriving these two heuristics for RRT* by locally linearizing the domain dynamics and applying linear quadratic regulation (LQR). The resulting algorithm, LQR-RRT*, finds optimal plans in domains with complex or underactuated dynamics without requiring domain-specific design choices. We demonstrate its application in domains that are successively torque-limited, underactuated, and in belief space.

We present a novel method for segmenting demonstrations, recognizing repeated skills, and generalizing complex tasks from unstructured demonstrations. This method combines many of the advantages of recent automatic segmentation methods for learning from demonstration into a single principled, integrated framework. Specifically, we use the Beta Process Autoregressive Hidden Markov Model and Dynamic Movement Primitives to learn and generalize a multi-step task on the PR2 mobile manipulator and to demonstrate the potential of our framework to learn a large library of skills over time.

We describe a robot system that autonomously acquires skills through interaction with its environment. The robot learns to sequence the execution of a set of innate controllers to solve a task, extracts and retains components of that solution as portable skills, and then transfers those skills to reduce the time required to learn to solve a second task.

We describe the Fourier basis, a linear value function approximation scheme based on the Fourier series. We empirically demonstrate that it performs well compared to radial basis functions and the polynomial basis, the two most popular fixed bases for linear value function approximation, and is competitive with learned proto-value functions.

We show that the lambda-return target used in the TD(lambda) family of algorithms is the maximum likelihood estimator for a specific model of how the variance of an n-step return estimate increases with n. We introduce the gamma-return estimator, an alternative target based on a more accurate model of variance, which defines the TDgamma family of complex-backup temporal difference learning algorithms. We derive TDgamma, the gamma-return equivalent of the original TD(lambda) algorithm, which eliminates the lambda parameter but can only perform updates at the end of an episode and requires time and space proportional to the episode length. We then derive a second algorithm, TDgamma(C), with a capacity parameter C. TDgamma(C) requires C times more time and memory than TD(lambda) and is incremental and online. We show that TDgamma outperforms TD(lambda) for any setting of lambda on 4 out of 5 benchmark domains, and that TDgamma(C) performs as well as or better than TD_gamma for intermediate settings of C.

We introduce CST, an algorithm for constructing skill trees from demonstration trajectories in continuous reinforcement learning domains. CST uses a changepoint detection method to segment each trajectory into a skill chain by detecting a change of appropriate abstraction, or that a segment is too complex to model as a single skill. The skill chains from each trajectory are then merged to form a skill tree. We demonstrate that CST constructs an appropriate skill tree that can be further refined through learning in a challenging continuous domain, and that it can be used to segment demonstration trajectories on a mobile manipulator into chains of skills where each skill is assigned an appropriate abstraction.

We introduce skill chaining, a skill discovery method for reinforcement learning agents in continuous domains, that builds chains of skills leading to an end-of-task reward. We demonstrate experimentally that it creates skills that result in performance benefits in a challenging continuous domain.

Robot positioning is an important function of autonomous intelligent robots. However, the application of external forces to a robot can disrupt its normal operation and cause localisation errors. We present a novel approach for detecting external disturbances based on optic flow without the use of egomotion information. Even though this research moderately validates the efficacy of the model we argue that its application is plausible to a large number of robotic systems.

We introduce the use of learned shaping rewards in reinforcement learning tasks, where an agent uses prior experience on a sequence of tasks to learn a portable predictor that estimates intermediate rewards, resulting in accelerated learning in later tasks that are related but distinct. Such agents can be trained on a sequence of relatively easy tasks in order to develop a more informative measure of reward that can be transferred to improve performance on more difficult tasks without requiring a hand coded shaping function. We use a rod positioning task to show that this significantly improves performance even after a very brief training period.

We describe the first part of a study investigating the usefulness of high school language results as a predictor of success in first year computer science courses at a university where students have widely varying English language skills. Our results indicate that contrary to the generally accepted view that achievement in high school mathematics courses is the best individual predictor of success in undergraduate computer science, success in English at the first-language level in high school correlates better with actual performance. We discuss the implications of this for universities whose medium of teaching is English, operating in social contexts where many students are not native English speakers.