Sanjiban Choudhury

Anirudh Vemula, Yuda Song, Aarti Singh et al. · cmu

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.

Wenting Zhao, Justin T Chiu, Jena D. Hwang et al. · allen-ai, berkeley

Language technologies that accurately model the dynamics of events must perform commonsense reasoning. Existing work evaluating commonsense reasoning focuses on making inferences about common, everyday situations. To instead investigate the ability to model unusual, unexpected, and unlikely situations, we explore the task of uncommonsense abductive reasoning. Given a piece of context with an unexpected outcome, this task requires reasoning abductively to generate an explanation that makes the unexpected outcome more likely in the context. To this end, we curate and release a new English language corpus called UNcommonsense. We characterize the performance differences between human explainers and the best-performing large language models, finding that model-enhanced human-written explanations achieve the highest quality by trading off between specificity and diversity. Finally, we experiment with several imitation learning algorithms to train open and accessible language models on this task. When compared with the vanilla supervised fine-tuning approach, these methods consistently reduce lose rates on both common and uncommonsense abductive reasoning judged by human evaluators.

Kushal Kedia, Prithwish Dan, Sanjiban Choudhury

Planning safe robot motions in the presence of humans requires reliable forecasts of future human motion. However, simply predicting the most likely motion from prior interactions does not guarantee safety. Such forecasts fail to model the long tail of possible events, which are rarely observed in limited datasets. On the other hand, planning for worst-case motions leads to overtly conservative behavior and a "frozen robot". Instead, we aim to learn forecasts that predict counterfactuals that humans guard against. We propose a novel game-theoretic framework for joint planning and forecasting with the payoff being the performance of the planner against the demonstrator, and present practical algorithms to train models in an end-to-end fashion. We demonstrate that our proposed algorithm results in safer plans in a crowd navigation simulator and real-world datasets of pedestrian motion. We release our code at https://github.com/portal-cornell/Game-Theoretic-Forecasting-Planning.

Gokul Swamy, Nived Rajaraman, Matthew Peng et al.

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 infinite sample regime, exact moment matching achieves value equivalence to the expert policy. However, in the finite sample regime, even if one has no optimization error, empirical variance can lead to a performance gap that scales with $H^2 / N$ for behavioral cloning and $H / \sqrt{N}$ for online moment matching, where $H$ is the horizon and $N$ 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 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 $\widetilde{O} \left( \min({H^{3/2}} / {N}, {H} / {\sqrt{N}} \right)$ 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.

Gokul Swamy, Sanjiban Choudhury, J. Andrew Bagnell et al.

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.

Gokul Swamy, Sanjiban Choudhury, J. Andrew Bagnell et al.

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.

Kushal Kedia, Atiksh Bhardwaj, Prithwish Dan et al.

In collaborative human-robot manipulation, a robot must predict human intents and adapt its actions accordingly to smoothly execute tasks. However, the human's intent in turn depends on actions the robot takes, creating a chicken-or-egg problem. Prior methods ignore such inter-dependency and instead train marginal intent prediction models independent of robot actions. This is because training conditional models is hard given a lack of paired human-robot interaction datasets. Can we instead leverage large-scale human-human interaction data that is more easily accessible? Our key insight is to exploit a correspondence between human and robot actions that enables transfer learning from human-human to human-robot data. We propose a novel architecture, InteRACT, that pre-trains a conditional intent prediction model on large human-human datasets and fine-tunes on a small human-robot dataset. We evaluate on a set of real-world collaborative human-robot manipulation tasks and show that our conditional model improves over various marginal baselines. We also introduce new techniques to tele-operate a 7-DoF robot arm and collect a diverse range of human-robot collaborative manipulation data, which we open-source.

David Durst, Feng Xie, Vishnu Sarukkai et al.

In multiplayer, first-person shooter games like Counter-Strike: Global Offensive (CS:GO), coordinated movement is a critical component of high-level strategic play. However, the complexity of team coordination and the variety of conditions present in popular game maps make it impractical to author hand-crafted movement policies for every scenario. We show that it is possible to take a data-driven approach to creating human-like movement controllers for CS:GO. We curate a team movement dataset comprising 123 hours of professional game play traces, and use this dataset to train a transformer-based movement model that generates human-like team movement for all players in a "Retakes" round of the game. Importantly, the movement prediction model is efficient. Performing inference for all players takes less than 0.5 ms per game step (amortized cost) on a single CPU core, making it plausible for use in commercial games today. Human evaluators assess that our model behaves more like humans than both commercially-available bots and procedural movement controllers scripted by experts (16% to 59% higher by TrueSkill rating of "human-like"). Using experiments involving in-game bot vs. bot self-play, we demonstrate that our model performs simple forms of teamwork, makes fewer common movement mistakes, and yields movement distributions, player lifetimes, and kill locations similar to those observed in professional CS:GO match play.

Gokul Swamy, Sanjiban Choudhury, J. Andrew Bagnell et al.

A variety of problems in econometrics and machine learning, including instrumental variable regression and Bellman residual minimization, can be formulated as satisfying a set of conditional moment restrictions (CMR). We derive a general, game-theoretic strategy for satisfying CMR that scales to nonlinear problems, is amenable to gradient-based optimization, and is able to account for finite sample uncertainty. We recover the approaches of Dikkala et al. and Dai et al. as special cases of our general framework before detailing various extensions and how to efficiently solve the game defined by CMR.

Kushal Kedia, Prithwish Dan, Angela Chao et al.

Human demonstrations as prompts are a powerful way to program robots to do long-horizon manipulation tasks. However, translating these demonstrations into robot-executable actions presents significant challenges due to execution mismatches in movement styles and physical capabilities. Existing methods for human-robot translation either depend on paired data, which is infeasible to scale, or rely heavily on frame-level visual similarities that often break down in practice. To address these challenges, we propose RHyME, a novel framework that automatically pairs human and robot trajectories using sequence-level optimal transport cost functions. Given long-horizon robot demonstrations, RHyME synthesizes semantically equivalent human videos by retrieving and composing short-horizon human clips. This approach facilitates effective policy training without the need for paired data. RHyME successfully imitates a range of cross-embodiment demonstrators, both in simulation and with a real human hand, achieving over 50% increase in task success compared to previous methods. We release our code and datasets at https://portal-cornell.github.io/rhyme/.

Sanjiban Choudhury

We introduce Agent Process Reward Models (AgentPRM), a simple and scalable framework for training LLM agents to continually improve through interactions. AgentPRM follows a lightweight actor-critic paradigm, using Monte Carlo rollouts to compute reward targets and optimize policies. It requires minimal modifications to existing RLHF pipelines, making it easy to integrate at scale. Beyond AgentPRM, we propose InversePRM, which learns process rewards directly from demonstrations without explicit outcome supervision. We also explore key challenges and opportunities, including exploration, process reward shaping, and model-predictive reasoning. We evaluate on ALFWorld benchmark, show that small 3B models trained with AgentPRM and InversePRM outperform strong GPT-4o baselines, and analyze test-time scaling, reward hacking, and more. Our code is available at: https://github.com/sanjibanc/agent_prm.

Kushal Kedia, Prithwish Dan, Atiksh Bhardwaj et al.

Seamless human-robot manipulation in close proximity relies on accurate forecasts of human motion. While there has been significant progress in learning forecast models at scale, when applied to manipulation tasks, these models accrue high errors at critical transition points leading to degradation in downstream planning performance. Our key insight is that instead of predicting the most likely human motion, it is sufficient to produce forecasts that capture how future human motion would affect the cost of a robot's plan. We present ManiCast, a novel framework that learns cost-aware human forecasts and feeds them to a model predictive control planner to execute collaborative manipulation tasks. Our framework enables fluid, real-time interactions between a human and a 7-DoF robot arm across a number of real-world tasks such as reactive stirring, object handovers, and collaborative table setting. We evaluate both the motion forecasts and the end-to-end forecaster-planner system against a range of learned and heuristic baselines while additionally contributing new datasets. We release our code and datasets at https://portal-cornell.github.io/manicast/.

Gonzalo Gonzalez-Pumariega, Leong Su Yean, Neha Sunkara et al.

Effective asynchronous planning, or the ability to efficiently reason and plan over states and actions that must happen in parallel or sequentially, is essential for agents that must account for time delays, reason over diverse long-horizon tasks, and collaborate with other agents. While large language model (LLM) agents show promise in high-level task planning, current benchmarks focus primarily on short-horizon tasks and do not evaluate such asynchronous planning capabilities. We introduce Robotouille, a challenging benchmark environment designed to test LLM agents' ability to handle long-horizon asynchronous scenarios. Our synchronous and asynchronous datasets capture increasingly complex planning challenges that go beyond existing benchmarks, requiring agents to manage overlapping tasks and interruptions. Our results show that ReAct (gpt4-o) achieves 47% on synchronous tasks but only 11% on asynchronous tasks, highlighting significant room for improvement. We further analyze failure modes, demonstrating the need for LLM agents to better incorporate long-horizon feedback and self-audit their reasoning during task execution. Code is available at https://github.com/portal-cornell/robotouille.

Arnav Kumar Jain, Gonzalo Gonzalez-Pumariega, Wayne Chen et al.

We address the problem of code generation from multi-turn execution feedback. Existing methods either generate code without feedback or use complex, hierarchical reinforcement learning to optimize multi-turn rewards. We propose a simple yet scalable approach, $μ$Code, that solves multi-turn code generation using only single-step rewards. Our key insight is that code generation is a one-step recoverable MDP, where the correct code can be recovered from any intermediate code state in a single turn. $μ$Code iteratively trains both a generator to provide code solutions conditioned on multi-turn execution feedback and a verifier to score the newly generated code. Experimental evaluations show that our approach achieves significant improvements over the state-of-the-art baselines. We provide analysis of the design choices of the reward models and policy, and show the efficacy of $μ$Code at utilizing the execution feedback. Our code is available at https://github.com/portal-cornell/muCode.

Arnav Kumar Jain, Vibhakar Mohta, Subin Kim et al.

The fundamental limitation of the behavioral cloning (BC) approach to imitation learning is that it only teaches an agent what the expert did at states the expert visited. This means that when a BC agent makes a mistake which takes them out of the support of the demonstrations, they often don't know how to recover from it. In this sense, BC is akin to giving the agent the fish -- giving them dense supervision across a narrow set of states -- rather than teaching them to fish: to be able to reason independently about achieving the expert's outcome even when faced with unseen situations at test-time. In response, we explore learning to search (L2S) from expert demonstrations, i.e. learning the components required to, at test time, plan to match expert outcomes, even after making a mistake. These include (1) a world model and (2) a reward model. We carefully ablate the set of algorithmic and design decisions required to combine these and other components for stable and sample/interaction-efficient learning of recovery behavior without additional human corrections. Across a dozen visual manipulation tasks from three benchmarks, our approach SAILOR consistently out-performs state-of-the-art Diffusion Policies trained via BC on the same data. Furthermore, scaling up the amount of demonstrations used for BC by 5-10x still leaves a performance gap. We find that SAILOR can identify nuanced failures and is robust to reward hacking. Our code is available at https://github.com/arnavkj1995/SAILOR .

William Huey, Huaxiaoyue Wang, Anne Wu et al.

We examine the problem of learning sequential tasks from a single visual demonstration. A key challenge arises when demonstrations are temporally misaligned due to variations in timing, differences in embodiment, or inconsistencies in execution. Existing approaches treat imitation as a distribution-matching problem, aligning individual frames between the agent and the demonstration. However, we show that such frame-level matching fails to enforce temporal ordering or ensure consistent progress. Our key insight is that matching should instead be defined at the level of sequences. We propose that perfect matching occurs when one sequence successfully covers all the subgoals in the same order as the other sequence. We present ORCA (ORdered Coverage Alignment), a dense per-timestep reward function that measures the probability of the agent covering demonstration frames in the correct order. On temporally misaligned demonstrations, we show that agents trained with the ORCA reward achieve $4.5$x improvement ($0.11 \rightarrow 0.50$ average normalized returns) for Meta-world tasks and $6.6$x improvement ($6.55 \rightarrow 43.3$ average returns) for Humanoid-v4 tasks compared to the best frame-level matching algorithms. We also provide empirical analysis showing that ORCA is robust to varying levels of temporal misalignment. Our code is available at https://github.com/portal-cornell/orca/

Gonzalo Gonzalez-Pumariega, Wayne Chen, Kushal Kedia et al.

Planning in complex environments requires an agent to efficiently query a world model to find a feasible sequence of actions from start to goal. Recent work has shown that Large Language Models (LLMs), with their rich prior knowledge and reasoning capabilities, can potentially help with planning by searching over promising states and adapting to feedback from the world. In this paper, we propose and study two fundamentally competing frameworks that leverage LLMs for query-efficient planning. The first uses LLMs as a heuristic within a search-based planner to select promising nodes to expand and propose promising actions. The second uses LLMs as a generative planner to propose an entire sequence of actions from start to goal, query a world model, and adapt based on feedback. We show that while both approaches improve upon comparable baselines, using an LLM as a generative planner results in significantly fewer interactions. Our key finding is that the LLM as a planner can more rapidly adapt its planning strategies based on immediate feedback than LLM as a heuristic. We present evaluations and ablations on Robotouille and PDDL planning benchmarks and discuss connections to existing theory on query-efficient planning algorithms. Code is available at https://github.com/portal-cornell/llms-for-planning

David Wu, Sanjiban Choudhury

Aligning large language models (LLMs) to human preferences is challenging in domains where preference data is unavailable. We address the problem of learning reward models for such target domains by leveraging feedback collected from simpler source domains, where human preferences are easier to obtain. Our key insight is that, while domains may differ significantly, human preferences convey \emph{domain-agnostic} concepts that can be effectively captured by a reward model. We propose \method, a framework that trains domain-invariant reward models by optimizing a dual loss: a domain loss that minimizes the divergence between source and target distribution, and a source loss that optimizes preferences on the source domain. We show \method is a general approach that we evaluate and analyze across 4 distinct settings: (1) Cross-lingual transfer (accuracy: $0.621 \rightarrow 0.661$), (2) Clean-to-noisy (accuracy: $0.671 \rightarrow 0.703$), (3) Few-shot-to-full transfer (accuracy: $0.845 \rightarrow 0.920$), and (4) Simple-to-complex tasks transfer (correlation: $0.508 \rightarrow 0.556$). Our code, models and data are available at \url{https://github.com/portal-cornell/dial}.

Siddhartha S. Srinivasa, Patrick Lancaster, Johan Michalove et al.

We present MuSHR, the Multi-agent System for non-Holonomic Racing. MuSHR is a low-cost, open-source robotic racecar platform for education and research, developed by the Personal Robotics Lab in the Paul G. Allen School of Computer Science & Engineering at the University of Washington. MuSHR aspires to contribute towards democratizing the field of robotics as a low-cost platform that can be built and deployed by following detailed, open documentation and do-it-yourself tutorials. A set of demos and lab assignments developed for the Mobile Robots course at the University of Washington provide guided, hands-on experience with the platform, and milestones for further development. MuSHR is a valuable asset for academic research labs, robotics instructors, and robotics enthusiasts.

Juntao Ren, Priya Sundaresan, Dorsa Sadigh et al.

Teaching robots to autonomously complete everyday tasks remains a challenge. Imitation Learning (IL) is a powerful approach that imbues robots with skills via demonstrations, but is limited by the labor-intensive process of collecting teleoperated robot data. Human videos offer a scalable alternative, but it remains difficult to directly train IL policies from them due to the lack of robot action labels. To address this, we propose to represent actions as short-horizon 2D trajectories on an image. These actions, or motion tracks, capture the predicted direction of motion for either human hands or robot end-effectors. We instantiate an IL policy called Motion Track Policy (MT-pi) which receives image observations and outputs motion tracks as actions. By leveraging this unified, cross-embodiment action space, MT-pi completes tasks with high success given just minutes of human video and limited additional robot demonstrations. At test time, we predict motion tracks from two camera views, recovering 6DoF trajectories via multi-view synthesis. MT-pi achieves an average success rate of 86.5% across 4 real-world tasks, outperforming state-of-the-art IL baselines which do not leverage human data or our action space by 40%, and generalizes to scenarios seen only in human videos. Code and videos are available on our website https://portal-cornell.github.io/motion_track_policy/.

Gokul Swamy, Sanjiban Choudhury, Wen Sun et al.

From a first-principles perspective, it may seem odd that the strongest results in foundation model fine-tuning (FT) are achieved via a relatively complex, two-stage training procedure. Specifically, one first trains a reward model (RM) on some dataset (e.g., human preferences) before using it to provide online feedback as part of a downstream reinforcement learning (RL) procedure, rather than directly optimizing the policy parameters on said dataset via offline maximum likelihood estimation. In fact, from an information-theoretic perspective, we can only lose information via passing through a reward model and cannot create any new information via on-policy sampling. To explain this discrepancy, we scrutinize several hypotheses on the value of RL in FT through both theoretical and empirical lenses. Of the hypotheses considered, we find the most support for the explanation that on problems with a generation-verification gap, (1) it is relatively easy to learn the relatively simple RM (verifier) from the preference data. Then, (2) the downstream RL procedure only returns policies (generators) that are optimal for such relatively simple verifiers. Thus, end-to-end, two-stage online FT only has to search over a reduced subset of the full space of policies, requiring less data than offline FT.

Juntao Ren, Gokul Swamy, Zhiwei Steven Wu et al.

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.

Arnav Kumar Jain, Harley Wiltzer, Jesse Farebrother et al.

In inverse reinforcement learning (IRL), an agent seeks to replicate expert demonstrations through interactions with the environment. Traditionally, IRL is treated as an adversarial game, where an adversary searches over reward models, and a learner optimizes the reward through repeated RL procedures. This game-solving approach is both computationally expensive and difficult to stabilize. In this work, we propose a novel approach to IRL by direct policy optimization: exploiting a linear factorization of the return as the inner product of successor features and a reward vector, we design an IRL algorithm by policy gradient descent on the gap between the learner and expert features. Our non-adversarial method does not require learning a reward function and can be solved seamlessly with existing actor-critic RL algorithms. Remarkably, our approach works in state-only settings without expert action labels, a setting which behavior cloning (BC) cannot solve. Empirical results demonstrate that our method learns from as few as a single expert demonstration and achieves improved performance on various control tasks.

Prithwish Dan, Kushal Kedia, Angela Chao et al.

Human videos offer a scalable way to train robot manipulation policies, but lack the action labels needed by standard imitation learning algorithms. Existing cross-embodiment approaches try to map human motion to robot actions, but often fail when the embodiments differ significantly. We propose X-Sim, a real-to-sim-to-real framework that uses object motion as a dense and transferable signal for learning robot policies. X-Sim starts by reconstructing a photorealistic simulation from an RGBD human video and tracking object trajectories to define object-centric rewards. These rewards are used to train a reinforcement learning (RL) policy in simulation. The learned policy is then distilled into an image-conditioned diffusion policy using synthetic rollouts rendered with varied viewpoints and lighting. To transfer to the real world, X-Sim introduces an online domain adaptation technique that aligns real and simulated observations during deployment. Importantly, X-Sim does not require any robot teleoperation data. We evaluate it across 5 manipulation tasks in 2 environments and show that it: (1) improves task progress by 30% on average over hand-tracking and sim-to-real baselines, (2) matches behavior cloning with 10x less data collection time, and (3) generalizes to new camera viewpoints and test-time changes. Code and videos are available at https://portal-cornell.github.io/X-Sim/.

Nicolas Espinosa-Dice, Sanjiban Choudhury, Wen Sun et al.

We consider the problem of imitation learning under misspecification: settings where the learner is fundamentally unable to replicate expert behavior everywhere. This is often true in practice due to differences in observation space and action space expressiveness (e.g. perceptual or morphological differences between robots and humans). Given the learner must make some mistakes in the misspecified setting, interaction with the environment is fundamentally required to figure out which mistakes are particularly costly and lead to compounding errors. However, given the computational cost and safety concerns inherent in interaction, we'd like to perform as little of it as possible while ensuring we've learned a strong policy. Accordingly, prior work has proposed a flavor of efficient inverse reinforcement learning algorithms that merely perform a computationally efficient local search procedure with strong guarantees in the realizable setting. We first prove that under a novel structural condition we term reward-agnostic policy completeness, these sorts of local-search based IRL algorithms are able to avoid compounding errors. We then consider the question of where we should perform local search in the first place, given the learner may not be able to "walk on a tightrope" as well as the expert in the misspecified setting. We prove that in the misspecified setting, it is beneficial to broaden the set of states on which local search is performed to include those reachable by good policies the learner can actually play. We then experimentally explore a variety of sources of misspecification and how offline data can be used to effectively broaden where we perform local search from.

Yujin Kim, Nathaniel Chin, Arnav Vasudev et al.

We study policy distillation under privileged information, where a student policy with only partial observations must learn from a teacher with full-state access. A key challenge is information asymmetry: the student cannot directly access the teacher's state space, leading to distributional shifts and policy degradation. Existing approaches either modify the teacher to produce realizable but sub-optimal demonstrations or rely on the student to explore missing information independently, both of which are inefficient. Our key insight is that the student should strategically interact with the teacher --querying only when necessary and resetting from recovery states --to stay on a recoverable path within its own observation space. We introduce two methods: (i) an imitation learning approach that adaptively determines when the student should query the teacher for corrections, and (ii) a reinforcement learning approach that selects where to initialize training for efficient exploration. We validate our methods in both simulated and real-world robotic tasks, demonstrating significant improvements over standard teacher-student baselines in training efficiency and final performance. The project website is available at : https://portal-cornell.github.io/CritiQ_ReTRy/

Rushit N. Shah, Nikolaos Agadakos, Synthia Sasulski et al.

Imitation learning often assumes that demonstrations are close to optimal according to some fixed, but unknown, cost function. However, according to satisficing theory, humans often choose acceptable behavior based on their personal (and potentially dynamic) levels of aspiration, rather than achieving (near-) optimality. For example, a lunar lander demonstration that successfully lands without crashing might be acceptable to a novice despite being slow or jerky. Using a margin-based objective to guide deep reinforcement learning, our focused satisficing approach to imitation learning seeks a policy that surpasses the demonstrator's aspiration levels -- defined over trajectories or portions of trajectories -- on unseen demonstrations without explicitly learning those aspirations. We show experimentally that this focuses the policy to imitate the highest quality (portions of) demonstrations better than existing imitation learning methods, providing much higher rates of guaranteed acceptability to the demonstrator, and competitive true returns on a range of environments.

David Wu, Gokul Swamy, J. Andrew Bagnell et al.

Inverse Reinforcement Learning (IRL) is a powerful framework for learning complex behaviors from expert demonstrations. However, it traditionally requires repeatedly solving a computationally expensive reinforcement learning (RL) problem in its inner loop. It is desirable to reduce the exploration burden by leveraging expert demonstrations in the inner-loop RL. As an example, recent work resets the learner to expert states in order to inform the learner of high-reward expert states. However, such an approach is infeasible in the real world. In this work, we consider an alternative approach to speeding up the RL subroutine in IRL: \emph{pessimism}, i.e., staying close to the expert's data distribution, instantiated via the use of offline RL algorithms. We formalize a connection between offline RL and IRL, enabling us to use an arbitrary offline RL algorithm to improve the sample efficiency of IRL. We validate our theory experimentally by demonstrating a strong correlation between the efficacy of an offline RL algorithm and how well it works as part of an IRL procedure. By using a strong offline RL algorithm as part of an IRL procedure, we are able to find policies that match expert performance significantly more efficiently than the prior art.

David Wu, Sanjiban Choudhury

Existing inverse reinforcement learning methods (e.g. MaxEntIRL, $f$-IRL) search over candidate reward functions and solve a reinforcement learning problem in the inner loop. This creates a rather strange inversion where a harder problem, reinforcement learning, is in the inner loop of a presumably easier problem, imitation learning. In this work, we show that better utilization of expert demonstrations can reduce the need for hard exploration in the inner RL loop, hence accelerating learning. Specifically, we propose two simple recipes: (1) placing expert transitions into the replay buffer of the inner RL algorithm (e.g. Soft-Actor Critic) which directly informs the learner about high reward states instead of forcing the learner to discover them through extensive exploration, and (2) using expert actions in Q value bootstrapping in order to improve the target Q value estimates and more accurately describe high value expert states. Our methods show significant gains over a MaxEntIRL baseline on the benchmark MuJoCo suite of tasks, speeding up recovery to 70\% of deterministic expert performance by 2.13x on HalfCheetah-v2, 2.6x on Ant-v2, 18x on Hopper-v2, and 3.36x on Walker2d-v2.

Konwoo Kim, Gokul Swamy, Zuxin Liu et al.

Regardless of the particular task we want them to perform in an environment, there are often shared safety constraints we want our agents to respect. For example, regardless of whether it is making a sandwich or clearing the table, a kitchen robot should not break a plate. Manually specifying such a constraint can be both time-consuming and error-prone. We show how to learn constraints from expert demonstrations of safe task completion by extending inverse reinforcement learning (IRL) techniques to the space of constraints. Intuitively, we learn constraints that forbid highly rewarding behavior that the expert could have taken but chose not to. Unfortunately, the constraint learning problem is rather ill-posed and typically leads to overly conservative constraints that forbid all behavior that the expert did not take. We counter this by leveraging diverse demonstrations that naturally occur in multi-task settings to learn a tighter set of constraints. We validate our method with simulation experiments on high-dimensional continuous control tasks.

Gokul Swamy, Sanjiban Choudhury, J. Andrew Bagnell et al.

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.

Mohak Bhardwaj, Sanjiban Choudhury, Byron Boots et al.

Lazy graph search algorithms are efficient at solving motion planning problems where edge evaluation is the computational bottleneck. These algorithms work by lazily computing the shortest potentially feasible path, evaluating edges along that path, and repeating until a feasible path is found. The order in which edges are selected is critical to minimizing the total number of edge evaluations: a good edge selector chooses edges that are not only likely to be invalid, but also eliminates future paths from consideration. We wish to learn such a selector by leveraging prior experience. We formulate this problem as a Markov Decision Process (MDP) on the state of the search problem. While solving this large MDP is generally intractable, we show that we can compute oracular selectors that can solve the MDP during training. With access to such oracles, we use imitation learning to find effective policies. If new search problems are sufficiently similar to problems solved during training, the learned policy will choose a good edge evaluation ordering and solve the motion planning problem quickly. We evaluate our algorithms on a wide range of 2D and 7D problems and show that the learned selector outperforms baseline commonly used heuristics. We further provide a novel theoretical analysis of lazy search in a Bayesian framework as well as regret guarantees on our imitation learning based approach to motion planning.

Gokul Swamy, Sanjiban Choudhury, J. Andrew Bagnell et al.

Recent work by Jarrett et al. attempts to frame the problem of offline imitation learning (IL) as one of learning a joint energy-based model, with the hope of out-performing standard behavioral cloning. We suggest that notational issues obscure how the psuedo-state visitation distribution the authors propose to optimize might be disconnected from the policy's $\textit{true}$ state visitation distribution. We further construct natural examples where the parameter coupling advocated by Jarrett et al. leads to inconsistent estimates of the expert's policy, unlike behavioral cloning.

Aditya Mandalika, Rosario Scalise, Brian Hou et al.

Sampling-based motion planners rely on incremental densification to discover progressively shorter paths. After computing feasible path $ξ$ between start $x_s$ and goal $x_t$, the Informed Set (IS) prunes the configuration space $\mathcal{C}$ by conservatively eliminating points that cannot yield shorter paths. Densification via sampling from this Informed Set retains asymptotic optimality of sampling from the entire configuration space. For path length $c(ξ)$ and Euclidean heuristic $h$, $IS = \{ x | x \in \mathcal{C}, h(x_s, x) + h(x, x_t) \leq c(ξ) \}$. Relying on the heuristic can render the IS especially conservative in high dimensions or complex environments. Furthermore, the IS only shrinks when shorter paths are discovered. Thus, the computational effort from each iteration of densification and planning is wasted if it fails to yield a shorter path, despite improving the cost-to-come for vertices in the search tree. Our key insight is that even in such a failure, shorter paths to vertices in the search tree (rather than just the goal) can immediately improve the planner's sampling strategy. Guided Incremental Local Densification (GuILD) leverages this information to sample from Local Subsets of the IS. We show that GuILD significantly outperforms uniform sampling of the Informed Set in simulated $\mathbb{R}^2$, $SE(2)$ environments and manipulation tasks in $\mathbb{R}^7$.

Matthew Schmittle, Sanjiban Choudhury, Siddhartha S. Srinivasa

A key challenge in Imitation Learning (IL) is that optimal state actions demonstrations are difficult for the teacher to provide. For example in robotics, providing kinesthetic demonstrations on a robotic manipulator requires the teacher to control multiple degrees of freedom at once. The difficulty of requiring optimal state action demonstrations limits the space of problems where the teacher can provide quality feedback. As an alternative to state action demonstrations, the teacher can provide corrective feedback such as their preferences or rewards. Prior work has created algorithms designed to learn from specific types of noisy feedback, but across teachers and tasks different forms of feedback may be required. Instead we propose that in order to learn from a diversity of scenarios we need to learn from a variety of feedback. To learn from a variety of feedback we make the following insight: the teacher's cost function is latent and we can model a stream of feedback as a stream of loss functions. We then use any online learning algorithm to minimize the sum of these losses. With this insight we can learn from a diversity of feedback that is weakly correlated with the teacher's true cost function. We unify prior work into a general corrective feedback meta-algorithm and show that regardless of feedback we can obtain the same regret bounds. We demonstrate our approach by learning to perform a household navigation task on a robotic racecar platform. Our results show that our approach can learn quickly from a variety of noisy feedback.

Gokul Swamy, Sanjiban Choudhury, J. Andrew Bagnell et al.

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.

Jonathan Spencer, Sanjiban Choudhury, Arun Venkatraman et al.

Imitation learning practitioners have often noted that conditioning policies on previous actions leads to a dramatic divergence between "held out" error and performance of the learner in situ. Interactive approaches can provably address this divergence but require repeated querying of a demonstrator. Recent work identifies this divergence as stemming from a "causal confound" in predicting the current action, and seek to ablate causal aspects of current state using tools from causal inference. In this work, we argue instead that this divergence is simply another manifestation of covariate shift, exacerbated particularly by settings of feedback between decisions and input features. The learner often comes to rely on features that are strongly predictive of decisions, but are subject to strong covariate shift. Our work demonstrates a broad class of problems where this shift can be mitigated, both theoretically and practically, by taking advantage of a simulator but without any further querying of expert demonstration. We analyze existing benchmarks used to test imitation learning approaches and find that these benchmarks are realizable and simple and thus insufficient for capturing the harder regimes of error compounding seen in real-world decision making problems. We find, in a surprising contrast with previous literature, but consistent with our theory, that naive behavioral cloning provides excellent results. We detail the need for new standardized benchmarks that capture the phenomena seen in robotics problems.

Mohak Bhardwaj, Sanjiban Choudhury, Byron Boots

Model-Predictive Control (MPC) is a powerful tool for controlling complex, real-world systems that uses a model to make predictions about future behavior. For each state encountered, MPC solves an online optimization problem to choose a control action that will minimize future cost. This is a surprisingly effective strategy, but real-time performance requirements warrant the use of simple models. If the model is not sufficiently accurate, then the resulting controller can be biased, limiting performance. We present a framework for improving on MPC with model-free reinforcement learning (RL). The key insight is to view MPC as constructing a series of local Q-function approximations. We show that by using a parameter $λ$, similar to the trace decay parameter in TD($λ$), we can systematically trade-off learned value estimates against the local Q-function approximations. We present a theoretical analysis that shows how error from inaccurate models in MPC and value function estimation in RL can be balanced. We further propose an algorithm that changes $λ$ over time to reduce the dependence on MPC as our estimates of the value function improve, and test the efficacy our approach on challenging high-dimensional manipulation tasks with biased models in simulation. We demonstrate that our approach can obtain performance comparable with MPC with access to true dynamics even under severe model bias and is more sample efficient as compared to model-free RL.

Brian Hou, Sanjiban Choudhury, Gilwoo Lee et al.

Collision checking is a computational bottleneck in motion planning, requiring lazy algorithms that explicitly reason about when to perform this computation. Optimism in the face of collision uncertainty minimizes the number of checks before finding the shortest path. However, this may take a prohibitively long time to compute, with no other feasible paths discovered during this period. For many real-time applications, we instead demand strong anytime performance, defined as minimizing the cumulative lengths of the feasible paths yielded over time. We introduce Posterior Sampling for Motion Planning (PSMP), an anytime lazy motion planning algorithm that leverages learned posteriors on edge collisions to quickly discover an initial feasible path and progressively yield shorter paths. PSMP obtains an expected regret bound of $\tilde{O}(\sqrt{\mathcal{S} \mathcal{A} T})$ and outperforms comparative baselines on a set of 2D and 7D planning problems.

Gilwoo Lee, Brian Hou, Sanjiban Choudhury et al.

Informed and robust decision making in the face of uncertainty is critical for robots that perform physical tasks alongside people. We formulate this as Bayesian Reinforcement Learning over latent Markov Decision Processes (MDPs). While Bayes-optimality is theoretically the gold standard, existing algorithms do not scale well to continuous state and action spaces. Our proposal builds on the following insight: in the absence of uncertainty, each latent MDP is easier to solve. We first obtain an ensemble of experts, one for each latent MDP, and fuse their advice to compute a baseline policy. Next, we train a Bayesian residual policy to improve upon the ensemble's recommendation and learn to reduce uncertainty. Our algorithm, Bayesian Residual Policy Optimization (BRPO), imports the scalability of policy gradient methods and task-specific expert skills. BRPO significantly improves the ensemble of experts and drastically outperforms existing adaptive RL methods.

Rogerio Bonatti, Wenshan Wang, Cherie Ho et al.

Aerial cinematography is revolutionizing industries that require live and dynamic camera viewpoints such as entertainment, sports, and security. However, safely piloting a drone while filming a moving target in the presence of obstacles is immensely taxing, often requiring multiple expert human operators. Hence, there is demand for an autonomous cinematographer that can reason about both geometry and scene context in real-time. Existing approaches do not address all aspects of this problem; they either require high-precision motion-capture systems or GPS tags to localize targets, rely on prior maps of the environment, plan for short time horizons, or only follow artistic guidelines specified before flight. In this work, we address the problem in its entirety and propose a complete system for real-time aerial cinematography that for the first time combines: (1) vision-based target estimation; (2) 3D signed-distance mapping for occlusion estimation; (3) efficient trajectory optimization for long time-horizon camera motion; and (4) learning-based artistic shot selection. We extensively evaluate our system both in simulation and in field experiments by filming dynamic targets moving through unstructured environments. Our results indicate that our system can operate reliably in the real world without restrictive assumptions. We also provide in-depth analysis and discussions for each module, with the hope that our design tradeoffs can generalize to other related applications. Videos of the complete system can be found at: https://youtu.be/ookhHnqmlaU.

Rahul Kumar, Aditya Mandalika, Sanjiban Choudhury et al.

We consider the problem of leveraging prior experience to generate roadmaps in sampling-based motion planning. A desirable roadmap is one that is sparse, allowing for fast search, with nodes spread out at key locations such that a low-cost feasible path exists. An increasingly popular approach is to learn a distribution of nodes that would produce such a roadmap. State-of-the-art is to train a conditional variational auto-encoder (CVAE) on the prior dataset with the shortest paths as target input. While this is quite effective on many problems, we show it can fail in the face of complex obstacle configurations or mismatch between training and testing. We present an algorithm LEGO that addresses these issues by training the CVAE with target samples that satisfy two important criteria. Firstly, these samples belong only to bottleneck regions along near-optimal paths that are otherwise difficult-to-sample with a uniform sampler. Secondly, these samples are spread out across diverse regions to maximize the likelihood of a feasible path existing. We formally define these properties and prove performance guarantees for LEGO. We extensively evaluate LEGO on a range of planning problems, including robot arm planning, and report significant gains over heuristics as well as learned baselines.

Mohak Bhardwaj, Sanjiban Choudhury, Byron Boots et al.

Lazy graph search algorithms are efficient at solving motion planning problems where edge evaluation is the computational bottleneck. These algorithms work by lazily computing the shortest potentially feasible path, evaluating edges along that path, and repeating until a feasible path is found. The order in which edges are selected is critical to minimizing the total number of edge evaluations: a good edge selector chooses edges that are not only likely to be invalid, but also eliminates future paths from consideration. We wish to learn such a selector by leveraging prior experience. We formulate this problem as a Markov Decision Process (MDP) on the state of the search problem. While solving this large MDP is generally intractable, we show that we can compute oracular selectors that can solve the MDP during training. With access to such oracles, we use imitation learning to find effective policies. If new search problems are sufficiently similar to problems solved during training, the learned policy will choose a good edge evaluation ordering and solve the motion planning problem quickly. We evaluate our algorithms on a wide range of 2D and 7D problems and show that the learned selector outperforms baseline commonly used heuristics.

Liyiming Ke, Sanjiban Choudhury, Matt Barnes et al.

We address the problem of imitation learning with multi-modal demonstrations. Instead of attempting to learn all modes, we argue that in many tasks it is sufficient to imitate any one of them. We show that the state-of-the-art methods such as GAIL and behavior cloning, due to their choice of loss function, often incorrectly interpolate between such modes. Our key insight is to minimize the right divergence between the learner and the expert state-action distributions, namely the reverse KL divergence or I-projection. We propose a general imitation learning framework for estimating and minimizing any f-Divergence. By plugging in different divergences, we are able to recover existing algorithms such as Behavior Cloning (Kullback-Leibler), GAIL (Jensen Shannon) and Dagger (Total Variation). Empirical results show that our approximate I-projection technique is able to imitate multi-modal behaviors more reliably than GAIL and behavior cloning.

Aditya Mandalika, Sanjiban Choudhury, Oren Salzman et al.

Lazy search algorithms can efficiently solve problems where edge evaluation is the bottleneck in computation, as is the case for robotic motion planning. The optimal algorithm in this class, LazySP, lazily restricts edge evaluation to only the shortest path. Doing so comes at the expense of search effort, i.e., LazySP must recompute the search tree every time an edge is found to be invalid. This becomes prohibitively expensive when dealing with large graphs or highly cluttered environments. Our key insight is the need to balance both edge evaluation and search effort to minimize the total planning time. Our contribution is two-fold. First, we propose a framework, Generalized Lazy Search (GLS), that seamlessly toggles between search and evaluation to prevent wasted efforts. We show that for a choice of toggle, GLS is provably more efficient than LazySP. Second, we leverage prior experience of edge probabilities to derive GLS policies that minimize expected planning time. We show that GLS equipped with such priors significantly outperforms competitive baselines for many simulated environments in R2, SE(2) and 7-DoF manipulation.

Rogerio Bonatti, Cherie Ho, Wenshan Wang et al.

The use of drones for aerial cinematography has revolutionized several applications and industries that require live and dynamic camera viewpoints such as entertainment, sports, and security. However, safely controlling a drone while filming a moving target usually requires multiple expert human operators; hence the need for an autonomous cinematographer. Current approaches have severe real-life limitations such as requiring fully scripted scenes, high-precision motion-capture systems or GPS tags to localize targets, and prior maps of the environment to avoid obstacles and plan for occlusion. In this work, we overcome such limitations and propose a complete system for aerial cinematography that combines: (1) a vision-based algorithm for target localization; (2) a real-time incremental 3D signed-distance map algorithm for occlusion and safety computation; and (3) a real-time camera motion planner that optimizes smoothness, collisions, occlusions and artistic guidelines. We evaluate robustness and real-time performance in series of field experiments and simulations by tracking dynamic targets moving through unknown, unstructured environments. Finally, we verify that despite removing previous limitations, our system achieves state-of-the-art performance. Videos of the system in action can be seen at https://youtu.be/ZE9MnCVmumc

Gilwoo Lee, Sanjiban Choudhury, Brian Hou et al.

We present the first PAC optimal algorithm for Bayes-Adaptive Markov Decision Processes (BAMDPs) in continuous state and action spaces, to the best of our knowledge. The BAMDP framework elegantly addresses model uncertainty by incorporating Bayesian belief updates into long-term expected return. However, computing an exact optimal Bayesian policy is intractable. Our key insight is to compute a near-optimal value function by covering the continuous state-belief-action space with a finite set of representative samples and exploiting the Lipschitz continuity of the value function. We prove the near-optimality of our algorithm and analyze a number of schemes that boost the algorithm's efficiency. Finally, we empirically validate our approach on a number of discrete and continuous BAMDPs and show that the learned policy has consistently competitive performance against baseline approaches.

Gilwoo Lee, Brian Hou, Aditya Mandalika et al.

Addressing uncertainty is critical for autonomous systems to robustly adapt to the real world. We formulate the problem of model uncertainty as a continuous Bayes-Adaptive Markov Decision Process (BAMDP), where an agent maintains a posterior distribution over latent model parameters given a history of observations and maximizes its expected long-term reward with respect to this belief distribution. Our algorithm, Bayesian Policy Optimization, builds on recent policy optimization algorithms to learn a universal policy that navigates the exploration-exploitation trade-off to maximize the Bayesian value function. To address challenges from discretizing the continuous latent parameter space, we propose a new policy network architecture that encodes the belief distribution independently from the observable state. Our method significantly outperforms algorithms that address model uncertainty without explicitly reasoning about belief distributions and is competitive with state-of-the-art Partially Observable Markov Decision Process solvers.

Rogerio Bonatti, Yanfu Zhang, Sanjiban Choudhury et al.

Autonomous aerial cinematography has the potential to enable automatic capture of aesthetically pleasing videos without requiring human intervention, empowering individuals with the capability of high-end film studios. Current approaches either only handle off-line trajectory generation, or offer strategies that reason over short time horizons and simplistic representations for obstacles, which result in jerky movement and low real-life applicability. In this work we develop a method for aerial filming that is able to trade off shot smoothness, occlusion, and cinematography guidelines in a principled manner, even under noisy actor predictions. We present a novel algorithm for real-time covariant gradient descent that we use to efficiently find the desired trajectories by optimizing a set of cost functions. Experimental results show that our approach creates attractive shots, avoiding obstacles and occlusion 65 times over 1.25 hours of flight time, re-planning at 5 Hz with a 10 s time horizon. We robustly film human actors, cars and bicycles performing different motion among obstacles, using various shot types.

Sankalp Arora, Sanjiban Choudhury, Sebastian Scherer

Partially Observable Markov Decision Processes (POMDPs) offer an elegant framework to model sequential decision making in uncertain environments. Solving POMDPs online is an active area of research and given the size of real-world problems approximate solvers are used. Recently, a few approaches have been suggested for solving POMDPs by using MDP solvers in conjunction with imitation learning. MDP based POMDP solvers work well for some cases, while catastrophically failing for others. The main failure point of such solvers is the lack of motivation for MDP solvers to gain information, since under their assumption the environment is either already known as much as it can be or the uncertainty will disappear after the next step. However for solving POMDP problems gaining information can lead to efficient solutions. In this paper we derive a set of conditions where MDP based POMDP solvers are provably sub-optimal. We then use the well-known tiger problem to demonstrate such sub-optimality. We show that multi-resolution, budgeted information gathering cannot be addressed using MDP based POMDP solvers. The contribution of the paper helps identify the properties of a POMDP problem for which the use of MDP based POMDP solvers is inappropriate, enabling better design choices.