Peter Stone

Anthony Francis, Claudia Pérez-D'Arpino, Chengshu Li et al. · cmu, mit

A major challenge to deploying robots widely is navigation in human-populated environments, commonly referred to as social robot navigation. While the field of social navigation has advanced tremendously in recent years, the fair evaluation of algorithms that tackle social navigation remains hard because it involves not just robotic agents moving in static environments but also dynamic human agents and their perceptions of the appropriateness of robot behavior. In contrast, clear, repeatable, and accessible benchmarks have accelerated progress in fields like computer vision, natural language processing and traditional robot navigation by enabling researchers to fairly compare algorithms, revealing limitations of existing solutions and illuminating promising new directions. We believe the same approach can benefit social navigation. In this paper, we pave the road towards common, widely accessible, and repeatable benchmarking criteria to evaluate social robot navigation. Our contributions include (a) a definition of a socially navigating robot as one that respects the principles of safety, comfort, legibility, politeness, social competency, agent understanding, proactivity, and responsiveness to context, (b) guidelines for the use of metrics, development of scenarios, benchmarks, datasets, and simulators to evaluate social navigation, and (c) a design of a social navigation metrics framework to make it easier to compare results from different simulators, robots and datasets.

Bo Liu, Qiang Liu, Peter Stone

As intelligent agents become autonomous over longer periods of time, they may eventually become lifelong counterparts to specific people. If so, it may be common for a user to want the agent to master a task temporarily but later on to forget the task due to privacy concerns. However enabling an agent to \emph{forget privately} what the user specified without degrading the rest of the learned knowledge is a challenging problem. With the aim of addressing this challenge, this paper formalizes this continual learning and private unlearning (CLPU) problem. The paper further introduces a straightforward but exactly private solution, CLPU-DER++, as the first step towards solving the CLPU problem, along with a set of carefully designed benchmark problems to evaluate the effectiveness of the proposed solution. The code is available at https://github.com/Cranial-XIX/Continual-Learning-Private-Unlearning.

Zifan Xu, Bo Liu, Xuesu Xiao et al.

Deep reinforcement learning (RL) has brought many successes for autonomous robot navigation. However, there still exists important limitations that prevent real-world use of RL-based navigation systems. For example, most learning approaches lack safety guarantees; and learned navigation systems may not generalize well to unseen environments. Despite a variety of recent learning techniques to tackle these challenges in general, a lack of an open-source benchmark and reproducible learning methods specifically for autonomous navigation makes it difficult for roboticists to choose what learning methods to use for their mobile robots and for learning researchers to identify current shortcomings of general learning methods for autonomous navigation. In this paper, we identify four major desiderata of applying deep RL approaches for autonomous navigation: (D1) reasoning under uncertainty, (D2) safety, (D3) learning from limited trial-and-error data, and (D4) generalization to diverse and novel environments. Then, we explore four major classes of learning techniques with the purpose of achieving one or more of the four desiderata: memory-based neural network architectures (D1), safe RL (D2), model-based RL (D2, D3), and domain randomization (D4). By deploying these learning techniques in a new open-source large-scale navigation benchmark and real-world environments, we perform a comprehensive study aimed at establishing to what extent can these techniques achieve these desiderata for RL-based navigation systems.

W. Bradley Knox, Stephane Hatgis-Kessell, Serena Booth et al.

The utility of reinforcement learning is limited by the alignment of reward functions with the interests of human stakeholders. One promising method for alignment is to learn the reward function from human-generated preferences between pairs of trajectory segments, a type of reinforcement learning from human feedback (RLHF). These human preferences are typically assumed to be informed solely by partial return, the sum of rewards along each segment. We find this assumption to be flawed and propose modeling human preferences instead as informed by each segment's regret, a measure of a segment's deviation from optimal decision-making. Given infinitely many preferences generated according to regret, we prove that we can identify a reward function equivalent to the reward function that generated those preferences, and we prove that the previous partial return model lacks this identifiability property in multiple contexts. We empirically show that our proposed regret preference model outperforms the partial return preference model with finite training data in otherwise the same setting. Additionally, we find that our proposed regret preference model better predicts real human preferences and also learns reward functions from these preferences that lead to policies that are better human-aligned. Overall, this work establishes that the choice of preference model is impactful, and our proposed regret preference model provides an improvement upon a core assumption of recent research. We have open sourced our experimental code, the human preferences dataset we gathered, and our training and preference elicitation interfaces for gathering a such a dataset.

Bo Liu, Yihao Feng, Peter Stone et al. · apple-ml

One of the grand enduring goals of AI is to create generalist agents that can learn multiple different tasks from diverse data via multitask learning (MTL). However, in practice, applying gradient descent (GD) on the average loss across all tasks may yield poor multitask performance due to severe under-optimization of certain tasks. Previous approaches that manipulate task gradients for a more balanced loss decrease require storing and computing all task gradients ($\mathcal{O}(k)$ space and time where $k$ is the number of tasks), limiting their use in large-scale scenarios. In this work, we introduce Fast Adaptive Multitask Optimization FAMO, a dynamic weighting method that decreases task losses in a balanced way using $\mathcal{O}(1)$ space and time. We conduct an extensive set of experiments covering multi-task supervised and reinforcement learning problems. Our results indicate that FAMO achieves comparable or superior performance to state-of-the-art gradient manipulation techniques while offering significant improvements in space and computational efficiency. Code is available at \url{https://github.com/Cranial-XIX/FAMO}.

Bo Liu, Yuqian Jiang, Xiaohan Zhang et al.

Large language models (LLMs) have demonstrated remarkable zero-shot generalization abilities: state-of-the-art chatbots can provide plausible answers to many common questions that arise in daily life. However, so far, LLMs cannot reliably solve long-horizon planning problems. By contrast, classical planners, once a problem is given in a formatted way, can use efficient search algorithms to quickly identify correct, or even optimal, plans. In an effort to get the best of both worlds, this paper introduces LLM+P, the first framework that incorporates the strengths of classical planners into LLMs. LLM+P takes in a natural language description of a planning problem, then returns a correct (or optimal) plan for solving that problem in natural language. LLM+P does so by first converting the language description into a file written in the planning domain definition language (PDDL), then leveraging classical planners to quickly find a solution, and then translating the found solution back into natural language. Along with LLM+P, we define a diverse set of different benchmark problems taken from common planning scenarios. Via a comprehensive set of experiments on these benchmark problems, we find that LLM+P is able to provide optimal solutions for most problems, while LLMs fail to provide even feasible plans for most problems.\footnote{The code and results are publicly available at https://github.com/Cranial-XIX/llm-pddl.git.

Bo Liu, Yifeng Zhu, Chongkai Gao et al. · apple-ml

Lifelong learning offers a promising paradigm of building a generalist agent that learns and adapts over its lifespan. Unlike traditional lifelong learning problems in image and text domains, which primarily involve the transfer of declarative knowledge of entities and concepts, lifelong learning in decision-making (LLDM) also necessitates the transfer of procedural knowledge, such as actions and behaviors. To advance research in LLDM, we introduce LIBERO, a novel benchmark of lifelong learning for robot manipulation. Specifically, LIBERO highlights five key research topics in LLDM: 1) how to efficiently transfer declarative knowledge, procedural knowledge, or the mixture of both; 2) how to design effective policy architectures and 3) effective algorithms for LLDM; 4) the robustness of a lifelong learner with respect to task ordering; and 5) the effect of model pretraining for LLDM. We develop an extendible procedural generation pipeline that can in principle generate infinitely many tasks. For benchmarking purpose, we create four task suites (130 tasks in total) that we use to investigate the above-mentioned research topics. To support sample-efficient learning, we provide high-quality human-teleoperated demonstration data for all tasks. Our extensive experiments present several insightful or even unexpected discoveries: sequential finetuning outperforms existing lifelong learning methods in forward transfer, no single visual encoder architecture excels at all types of knowledge transfer, and naive supervised pretraining can hinder agents' performance in the subsequent LLDM. Check the website at https://libero-project.github.io for the code and the datasets.

Peter Stone, Rodney Brooks, Erik Brynjolfsson et al.

In September 2016, Stanford's "One Hundred Year Study on Artificial Intelligence" project (AI100) issued the first report of its planned long-term periodic assessment of artificial intelligence (AI) and its impact on society. It was written by a panel of 17 study authors, each of whom is deeply rooted in AI research, chaired by Peter Stone of the University of Texas at Austin. The report, entitled "Artificial Intelligence and Life in 2030," examines eight domains of typical urban settings on which AI is likely to have impact over the coming years: transportation, home and service robots, healthcare, education, public safety and security, low-resource communities, employment and workplace, and entertainment. It aims to provide the general public with a scientifically and technologically accurate portrayal of the current state of AI and its potential and to help guide decisions in industry and governments, as well as to inform research and development in the field. The charge for this report was given to the panel by the AI100 Standing Committee, chaired by Barbara Grosz of Harvard University.

Haresh Karnan, Anirudh Nair, Xuesu Xiao et al.

Social navigation is the capability of an autonomous agent, such as a robot, to navigate in a 'socially compliant' manner in the presence of other intelligent agents such as humans. With the emergence of autonomously navigating mobile robots in human populated environments (e.g., domestic service robots in homes and restaurants and food delivery robots on public sidewalks), incorporating socially compliant navigation behaviors on these robots becomes critical to ensuring safe and comfortable human robot coexistence. To address this challenge, imitation learning is a promising framework, since it is easier for humans to demonstrate the task of social navigation rather than to formulate reward functions that accurately capture the complex multi objective setting of social navigation. The use of imitation learning and inverse reinforcement learning to social navigation for mobile robots, however, is currently hindered by a lack of large scale datasets that capture socially compliant robot navigation demonstrations in the wild. To fill this gap, we introduce Socially CompliAnt Navigation Dataset (SCAND) a large scale, first person view dataset of socially compliant navigation demonstrations. Our dataset contains 8.7 hours, 138 trajectories, 25 miles of socially compliant, human teleoperated driving demonstrations that comprises multi modal data streams including 3D lidar, joystick commands, odometry, visual and inertial information, collected on two morphologically different mobile robots a Boston Dynamics Spot and a Clearpath Jackal by four different human demonstrators in both indoor and outdoor environments. We additionally perform preliminary analysis and validation through real world robot experiments and show that navigation policies learned by imitation learning on SCAND generate socially compliant behaviors

Jiaxun Cui, Hang Qiu, Dian Chen et al.

Optical sensors and learning algorithms for autonomous vehicles have dramatically advanced in the past few years. Nonetheless, the reliability of today's autonomous vehicles is hindered by the limited line-of-sight sensing capability and the brittleness of data-driven methods in handling extreme situations. With recent developments of telecommunication technologies, cooperative perception with vehicle-to-vehicle communications has become a promising paradigm to enhance autonomous driving in dangerous or emergency situations. We introduce COOPERNAUT, an end-to-end learning model that uses cross-vehicle perception for vision-based cooperative driving. Our model encodes LiDAR information into compact point-based representations that can be transmitted as messages between vehicles via realistic wireless channels. To evaluate our model, we develop AutoCastSim, a network-augmented driving simulation framework with example accident-prone scenarios. Our experiments on AutoCastSim suggest that our cooperative perception driving models lead to a 40% improvement in average success rate over egocentric driving models in these challenging driving situations and a 5 times smaller bandwidth requirement than prior work V2VNet. COOPERNAUT and AUTOCASTSIM are available at https://ut-austin-rpl.github.io/Coopernaut/.

Zizhao Wang, Xuesu Xiao, Zifan Xu et al.

Learning dynamics models accurately is an important goal for Model-Based Reinforcement Learning (MBRL), but most MBRL methods learn a dense dynamics model which is vulnerable to spurious correlations and therefore generalizes poorly to unseen states. In this paper, we introduce Causal Dynamics Learning for Task-Independent State Abstraction (CDL), which first learns a theoretically proved causal dynamics model that removes unnecessary dependencies between state variables and the action, thus generalizing well to unseen states. A state abstraction can then be derived from the learned dynamics, which not only improves sample efficiency but also applies to a wider range of tasks than existing state abstraction methods. Evaluated on two simulated environments and downstream tasks, both the dynamics model and policies learned by the proposed method generalize well to unseen states and the derived state abstraction improves sample efficiency compared to learning without it.

Haresh Karnan, Elvin Yang, Daniel Farkash et al.

Terrain awareness, i.e., the ability to identify and distinguish different types of terrain, is a critical ability that robots must have to succeed at autonomous off-road navigation. Current approaches that provide robots with this awareness either rely on labeled data which is expensive to collect, engineered features and cost functions that may not generalize, or expert human demonstrations which may not be available. Towards endowing robots with terrain awareness without these limitations, we introduce Self-supervised TErrain Representation LearnING (STERLING), a novel approach for learning terrain representations that relies solely on easy-to-collect, unconstrained (e.g., non-expert), and unlabelled robot experience, with no additional constraints on data collection. STERLING employs a novel multi-modal self-supervision objective through non-contrastive representation learning to learn relevant terrain representations for terrain-aware navigation. Through physical robot experiments in off-road environments, we evaluate STERLING features on the task of preference-aligned visual navigation and find that STERLING features perform on par with fully supervised approaches and outperform other state-of-the-art methods with respect to preference alignment. Additionally, we perform a large-scale experiment of autonomously hiking a 3-mile long trail which STERLING completes successfully with only two manual interventions, demonstrating its robustness to real-world off-road conditions.

Megan M. Baker, Alexander New, Mario Aguilar-Simon et al.

Despite the advancement of machine learning techniques in recent years, state-of-the-art systems lack robustness to "real world" events, where the input distributions and tasks encountered by the deployed systems will not be limited to the original training context, and systems will instead need to adapt to novel distributions and tasks while deployed. This critical gap may be addressed through the development of "Lifelong Learning" systems that are capable of 1) Continuous Learning, 2) Transfer and Adaptation, and 3) Scalability. Unfortunately, efforts to improve these capabilities are typically treated as distinct areas of research that are assessed independently, without regard to the impact of each separate capability on other aspects of the system. We instead propose a holistic approach, using a suite of metrics and an evaluation framework to assess Lifelong Learning in a principled way that is agnostic to specific domains or system techniques. Through five case studies, we show that this suite of metrics can inform the development of varied and complex Lifelong Learning systems. We highlight how the proposed suite of metrics quantifies performance trade-offs present during Lifelong Learning system development - both the widely discussed Stability-Plasticity dilemma and the newly proposed relationship between Sample Efficient and Robust Learning. Further, we make recommendations for the formulation and use of metrics to guide the continuing development of Lifelong Learning systems and assess their progress in the future.

Vaibhav Bajaj, Guni Sharon, Peter Stone

Applying reinforcement learning (RL) to sparse reward domains is notoriously challenging due to insufficient guiding signals. Common RL techniques for addressing such domains include (1) learning from demonstrations and (2) curriculum learning. While these two approaches have been studied in detail, they have rarely been considered together. This paper aims to do so by introducing a principled task phasing approach that uses demonstrations to automatically generate a curriculum sequence. Using inverse RL from (suboptimal) demonstrations we define a simple initial task. Our task phasing approach then provides a framework to gradually increase the complexity of the task all the way to the target task, while retuning the RL agent in each phasing iteration. Two approaches for phasing are considered: (1) gradually increasing the proportion of time steps an RL agent is in control, and (2) phasing out a guiding informative reward function. We present conditions that guarantee the convergence of these approaches to an optimal policy. Experimental results on 3 sparse reward domains demonstrate that our task phasing approaches outperform state-of-the-art approaches with respect to asymptotic performance.

Bo Liu, Rui Wang, Lemeng Wu et al. · apple-ml

Modern large language models are built on sequence modeling via next-token prediction. While the Transformer remains the dominant architecture for sequence modeling, its quadratic decoding complexity in sequence length poses a major limitation. State-space models (SSMs) present a competitive alternative, offering linear decoding efficiency while maintaining parallelism during training. However, most existing SSMs rely on linear recurrence designs that appear somewhat ad hoc. In this work, we explore SSM design through the lens of online learning, conceptualizing SSMs as meta-modules for specific online learning problems. This approach links SSM design to formulating precise online learning objectives, with state transition rules derived from solving these objectives. Based on this insight, we introduce a novel deep SSM architecture, Longhorn, whose update resembles the closed-form solution for solving the online associative recall problem. Our experimental results show that Longhorn outperforms state-of-the-art SSMs, including the Mamba model, on standard sequence modeling benchmarks, language modeling, and vision tasks. Specifically, Longhorn achieves a 1.8x improvement in sample efficiency compared to Mamba, and can extrapolate over contexts that are up to 16x longer during inference.

Jiaheng Hu, Rose Hendrix, Ali Farhadi et al.

In recent years, the Robotics field has initiated several efforts toward building generalist robot policies through large-scale multi-task Behavior Cloning. However, direct deployments of these policies have led to unsatisfactory performance, where the policy struggles with unseen states and tasks. How can we break through the performance plateau of these models and elevate their capabilities to new heights? In this paper, we propose FLaRe, a large-scale Reinforcement Learning fine-tuning framework that integrates robust pre-trained representations, large-scale training, and gradient stabilization techniques. Our method aligns pre-trained policies towards task completion, achieving state-of-the-art (SoTA) performance both on previously demonstrated and on entirely novel tasks and embodiments. Specifically, on a set of long-horizon mobile manipulation tasks, FLaRe achieves an average success rate of 79.5% in unseen environments, with absolute improvements of +23.6% in simulation and +30.7% on real robots over prior SoTA methods. By utilizing only sparse rewards, our approach can enable generalizing to new capabilities beyond the pretraining data with minimal human effort. Moreover, we demonstrate rapid adaptation to new embodiments and behaviors with less than a day of fine-tuning. Videos can be found on the project website at https://robot-flare.github.io/

Bo Liu, Yihao Feng, Qiang Liu et al. · apple-ml

Goal-conditioned reinforcement learning (GCRL) has a wide range of potential real-world applications, including manipulation and navigation problems in robotics. Especially in such robotics tasks, sample efficiency is of the utmost importance for GCRL since, by default, the agent is only rewarded when it reaches its goal. While several methods have been proposed to improve the sample efficiency of GCRL, one relatively under-studied approach is the design of neural architectures to support sample efficiency. In this work, we introduce a novel neural architecture for GCRL that achieves significantly better sample efficiency than the commonly-used monolithic network architecture. The key insight is that the optimal action-value function Q^*(s, a, g) must satisfy the triangle inequality in a specific sense. Furthermore, we introduce the metric residual network (MRN) that deliberately decomposes the action-value function Q(s,a,g) into the negated summation of a metric plus a residual asymmetric component. MRN provably approximates any optimal action-value function Q^*(s,a,g), thus making it a fitting neural architecture for GCRL. We conduct comprehensive experiments across 12 standard benchmark environments in GCRL. The empirical results demonstrate that MRN uniformly outperforms other state-of-the-art GCRL neural architectures in terms of sample efficiency.

Caroline Wang, Ishan Durugkar, Elad Liebman et al.

Current approaches to multi-agent cooperation rely heavily on centralized mechanisms or explicit communication protocols to ensure convergence. This paper studies the problem of distributed multi-agent learning without resorting to centralized components or explicit communication. It examines the use of distribution matching to facilitate the coordination of independent agents. In the proposed scheme, each agent independently minimizes the distribution mismatch to the corresponding component of a target visitation distribution. The theoretical analysis shows that under certain conditions, each agent minimizing its individual distribution mismatch allows the convergence to the joint policy that generated the target distribution. Further, if the target distribution is from a joint policy that optimizes a cooperative task, the optimal policy for a combination of this task reward and the distribution matching reward is the same joint policy. This insight is used to formulate a practical algorithm (DM$^2$), in which each individual agent matches a target distribution derived from concurrently sampled trajectories from a joint expert policy. Experimental validation on the StarCraft domain shows that combining (1) a task reward, and (2) a distribution matching reward for expert demonstrations for the same task, allows agents to outperform a naive distributed baseline. Additional experiments probe the conditions under which expert demonstrations need to be sampled to obtain the learning benefits.

Jiaheng Hu, Jay Shim, Chen Tang et al.

Continual Reinforcement Learning (CRL) for Vision-Language-Action (VLA) models is a promising direction toward self-improving embodied agents that can adapt in openended, evolving environments. However, conventional wisdom from continual learning suggests that naive Sequential Fine-Tuning (Seq. FT) leads to catastrophic forgetting, necessitating complex CRL strategies. In this work, we take a step back and conduct a systematic study of CRL for large pretrained VLAs across three models and five challenging lifelong RL benchmarks. We find that, contrary to established belief, simple Seq. FT with low-rank adaptation (LoRA) is remarkably strong: it achieves high plasticity, exhibits little to no forgetting, and retains strong zero-shot generalization, frequently outperforming more sophisticated CRL methods. Through detailed analysis, we show that this robustness arises from a synergy between the large pretrained model, parameter-efficient adaptation, and on-policy RL. Together, these components reshape the stability-plasticity trade-off, making continual adaptation both stable and scalable. Our results position Sequential Fine-Tuning as a powerful method for continual RL with VLAs and provide new insights into lifelong learning in the large model era. Code is available at github.com/UT-Austin-RobIn/continual-vla-rl.

Mao Ye, Bo Liu, Stephen Wright et al.

Bilevel optimization (BO) is useful for solving a variety of important machine learning problems including but not limited to hyperparameter optimization, meta-learning, continual learning, and reinforcement learning. Conventional BO methods need to differentiate through the low-level optimization process with implicit differentiation, which requires expensive calculations related to the Hessian matrix. There has been a recent quest for first-order methods for BO, but the methods proposed to date tend to be complicated and impractical for large-scale deep learning applications. In this work, we propose a simple first-order BO algorithm that depends only on first-order gradient information, requires no implicit differentiation, and is practical and efficient for large-scale non-convex functions in deep learning. We provide non-asymptotic convergence analysis of the proposed method to stationary points for non-convex objectives and present empirical results that show its superior practical performance.

Akarsh Kumar, Bo Liu, Risto Miikkulainen et al.

Evolutionary algorithms are sensitive to the mutation rate (MR); no single value of this parameter works well across domains. Self-adaptive MR approaches have been proposed but they tend to be brittle: Sometimes they decay the MR to zero, thus halting evolution. To make self-adaptive MR robust, this paper introduces the Group Elite Selection of Mutation Rates (GESMR) algorithm. GESMR co-evolves a population of solutions and a population of MRs, such that each MR is assigned to a group of solutions. The resulting best mutational change in the group, instead of average mutational change, is used for MR selection during evolution, thus avoiding the vanishing MR problem. With the same number of function evaluations and with almost no overhead, GESMR converges faster and to better solutions than previous approaches on a wide range of continuous test optimization problems. GESMR also scales well to high-dimensional neuroevolution for supervised image-classification tasks and for reinforcement learning control tasks. Remarkably, GESMR produces MRs that are optimal in the long-term, as demonstrated through a comprehensive look-ahead grid search. Thus, GESMR and its theoretical and empirical analysis demonstrate how self-adaptation can be harnessed to improve performance in several applications of evolutionary computation.

Caroline Wang, Garrett Warnell, Peter Stone

While combining imitation learning (IL) and reinforcement learning (RL) is a promising way to address poor sample efficiency in autonomous behavior acquisition, methods that do so typically assume that the requisite behavior demonstrations are provided by an expert that behaves optimally with respect to a task reward. If, however, suboptimal demonstrations are provided, a fundamental challenge appears in that the demonstration-matching objective of IL conflicts with the return-maximization objective of RL. This paper introduces D-Shape, a new method for combining IL and RL that uses ideas from reward shaping and goal-conditioned RL to resolve the above conflict. D-Shape allows learning from suboptimal demonstrations while retaining the ability to find the optimal policy with respect to the task reward. We experimentally validate D-Shape in sparse-reward gridworld domains, showing that it both improves over RL in terms of sample efficiency and converges consistently to the optimal policy in the presence of suboptimal demonstrations.

Elad Liebman, Peter Stone

Past research has clearly established that music can affect mood and that mood affects emotional and cognitive processing, and thus decision-making. It follows that if a robot interacting with a person needs to predict the person's behavior, knowledge of the music the person is listening to when acting is a potentially relevant feature. To date, however, there has not been any concrete evidence that a robot can improve its human-interactive decision-making by taking into account what the person is listening to. This research fills this gap by reporting the results of an experiment in which human participants were required to complete a task in the presence of an autonomous agent while listening to background music. Specifically, the participants drove a simulated car through an intersection while listening to music. The intersection was not empty, as another simulated vehicle, controlled autonomously, was also crossing the intersection in a different direction. Our results clearly indicate that such background information can be effectively incorporated in an agent's world representation in order to better predict people's behavior. We subsequently analyze how knowledge of music impacted both participant behavior and the resulting learned policy.\setcounter{footnote}{2}\footnote{An earlier version of part of the material in this paper appeared originally in the first author's Ph.D. Dissertation~\cite{liebman2020sequential} but it has not appeared in any pear-reviewed conference or journal.}

Eddy Hudson, Ishan Durugkar, Garrett Warnell et al.

Given a dataset of expert agent interactions with an environment of interest, a viable method to extract an effective agent policy is to estimate the maximum likelihood policy indicated by this data. This approach is commonly referred to as behavioral cloning (BC). In this work, we describe a key disadvantage of BC that arises due to the maximum likelihood objective function; namely that BC is mean-seeking with respect to the state-conditional expert action distribution when the learner's policy is represented with a Gaussian. To address this issue, we introduce a modified version of BC, Adversarial Behavioral Cloning (ABC), that exhibits mode-seeking behavior by incorporating elements of GAN (generative adversarial network) training. We evaluate ABC on toy domains and a domain based on Hopper from the DeepMind Control suite, and show that it outperforms standard BC by being mode-seeking in nature.

Chen Tang, Ben Abbatematteo, Jiaheng Hu et al.

Reinforcement learning (RL), particularly its combination with deep neural networks referred to as deep RL (DRL), has shown tremendous promise across a wide range of applications, suggesting its potential for enabling the development of sophisticated robotic behaviors. Robotics problems, however, pose fundamental difficulties for the application of RL, stemming from the complexity and cost of interacting with the physical world. This article provides a modern survey of DRL for robotics, with a particular focus on evaluating the real-world successes achieved with DRL in realizing several key robotic competencies. Our analysis aims to identify the key factors underlying those exciting successes, reveal underexplored areas, and provide an overall characterization of the status of DRL in robotics. We highlight several important avenues for future work, emphasizing the need for stable and sample-efficient real-world RL paradigms, holistic approaches for discovering and integrating various competencies to tackle complex long-horizon, open-world tasks, and principled development and evaluation procedures. This survey is designed to offer insights for both RL practitioners and roboticists toward harnessing RL's power to create generally capable real-world robotic systems.

W. Bradley Knox, Stephane Hatgis-Kessell, Sigurdur Orn Adalgeirsson et al.

We consider algorithms for learning reward functions from human preferences over pairs of trajectory segments, as used in reinforcement learning from human feedback (RLHF). Most recent work assumes that human preferences are generated based only upon the reward accrued within those segments, or their partial return. Recent work casts doubt on the validity of this assumption, proposing an alternative preference model based upon regret. We investigate the consequences of assuming preferences are based upon partial return when they actually arise from regret. We argue that the learned function is an approximation of the optimal advantage function, $\hat{A^*_r}$, not a reward function. We find that if a specific pitfall is addressed, this incorrect assumption is not particularly harmful, resulting in a highly shaped reward function. Nonetheless, this incorrect usage of $\hat{A^*_r}$ is less desirable than the appropriate and simpler approach of greedy maximization of $\hat{A^*_r}$. From the perspective of the regret preference model, we also provide a clearer interpretation of fine tuning contemporary large language models with RLHF. This paper overall provides insight regarding why learning under the partial return preference model tends to work so well in practice, despite it conforming poorly to how humans give preferences.

Chandler Smith, Marwa Abdulhai, Manfred Diaz et al.

Large Language Model (LLM) agents have demonstrated impressive capabilities for social interaction and are increasingly being deployed in situations where they might engage with both human and artificial agents. These interactions represent a critical frontier for LLM-based agents, yet existing evaluation methods fail to measure how well these capabilities generalize to novel social situations. In this paper, we introduce a method for evaluating the ability of LLM-based agents to cooperate in zero-shot, mixed-motive environments using Concordia, a natural language multi-agent simulation environment. Our method measures general cooperative intelligence by testing an agent's ability to identify and exploit opportunities for mutual gain across diverse partners and contexts. We present empirical results from the NeurIPS 2024 Concordia Contest, where agents were evaluated on their ability to achieve mutual gains across a suite of diverse scenarios ranging from negotiation to collective action problems. Our findings reveal significant gaps between current agent capabilities and the robust generalization required for reliable cooperation, particularly in scenarios demanding persuasion and norm enforcement.

Hager Radi, Josiah P. Hanna, Peter Stone et al.

The world currently offers an abundance of data in multiple domains, from which we can learn reinforcement learning (RL) policies without further interaction with the environment. RL agents learning offline from such data is possible but deploying them while learning might be dangerous in domains where safety is critical. Therefore, it is essential to find a way to estimate how a newly-learned agent will perform if deployed in the target environment before actually deploying it and without the risk of overestimating its true performance. To achieve this, we introduce a framework for safe evaluation of offline learning using approximate high-confidence off-policy evaluation (HCOPE) to estimate the performance of offline policies during learning. In our setting, we assume a source of data, which we split into a train-set, to learn an offline policy, and a test-set, to estimate a lower-bound on the offline policy using off-policy evaluation with bootstrapping. A lower-bound estimate tells us how good a newly-learned target policy would perform before it is deployed in the real environment, and therefore allows us to decide when to deploy our learned policy.

Zifan Xu, Myoungkyu Seo, Dongmyeong Lee et al.

Learning fast and robust ball-kicking skills is a critical capability for humanoid soccer robots, yet it remains a challenging problem due to the need for rapid leg swings, postural stability on a single support foot, and robustness under noisy sensory input and external perturbations (e.g., opponents). This paper presents a reinforcement learning (RL)-based system that enables humanoid robots to execute robust continual ball-kicking with adaptability to different ball-goal configurations. The system extends a typical teacher-student training framework -- in which a "teacher" policy is trained with ground truth state information and the "student" learns to mimic it with noisy, imperfect sensing -- by including four training stages: (1) long-distance ball chasing (teacher); (2) directional kicking (teacher); (3) teacher policy distillation (student); and (4) student adaptation and refinement (student). Key design elements -- including tailored reward functions, realistic noise modeling, and online constrained RL for adaptation and refinement -- are critical for closing the sim-to-real gap and sustaining performance under perceptual uncertainty. Extensive evaluations in both simulation and on a real robot demonstrate strong kicking accuracy and goal-scoring success across diverse ball-goal configurations. Ablation studies further highlight the necessity of the constrained RL, noise modeling, and the adaptation stage. This work presents a system for learning robust continual humanoid ball-kicking under imperfect perception, establishing a benchmark task for visuomotor skill learning in humanoid whole-body control.

Haresh Karnan, Elvin Yang, Garrett Warnell et al.

Autonomous mobility tasks such as lastmile delivery require reasoning about operator indicated preferences over terrains on which the robot should navigate to ensure both robot safety and mission success. However, coping with out of distribution data from novel terrains or appearance changes due to lighting variations remains a fundamental problem in visual terrain adaptive navigation. Existing solutions either require labor intensive manual data recollection and labeling or use handcoded reward functions that may not align with operator preferences. In this work, we posit that operator preferences for visually novel terrains, which the robot should adhere to, can often be extrapolated from established terrain references within the inertial, proprioceptive, and tactile domain. Leveraging this insight, we introduce Preference extrApolation for Terrain awarE Robot Navigation, PATERN, a novel framework for extrapolating operator terrain preferences for visual navigation. PATERN learns to map inertial, proprioceptive, tactile measurements from the robots observations to a representation space and performs nearest neighbor search in this space to estimate operator preferences over novel terrains. Through physical robot experiments in outdoor environments, we assess PATERNs capability to extrapolate preferences and generalize to novel terrains and challenging lighting conditions. Compared to baseline approaches, our findings indicate that PATERN robustly generalizes to diverse terrains and varied lighting conditions, while navigating in a preference aligned manner.

James MacGlashan, Evan Archer, Alisa Devlic et al.

Designing reinforcement learning (RL) agents is typically a difficult process that requires numerous design iterations. Learning can fail for a multitude of reasons, and standard RL methods provide too few tools to provide insight into the exact cause. In this paper, we show how to integrate value decomposition into a broad class of actor-critic algorithms and use it to assist in the iterative agent-design process. Value decomposition separates a reward function into distinct components and learns value estimates for each. These value estimates provide insight into an agent's learning and decision-making process and enable new training methods to mitigate common problems. As a demonstration, we introduce SAC-D, a variant of soft actor-critic (SAC) adapted for value decomposition. SAC-D maintains similar performance to SAC, while learning a larger set of value predictions. We also introduce decomposition-based tools that exploit this information, including a new reward influence metric, which measures each reward component's effect on agent decision-making. Using these tools, we provide several demonstrations of decomposition's use in identifying and addressing problems in the design of both environments and agents. Value decomposition is broadly applicable and easy to incorporate into existing algorithms and workflows, making it a powerful tool in an RL practitioner's toolbox.

Yifeng Zhu, Zhenyu Jiang, Peter Stone et al.

We introduce GROOT, an imitation learning method for learning robust policies with object-centric and 3D priors. GROOT builds policies that generalize beyond their initial training conditions for vision-based manipulation. It constructs object-centric 3D representations that are robust toward background changes and camera views and reason over these representations using a transformer-based policy. Furthermore, we introduce a segmentation correspondence model that allows policies to generalize to new objects at test time. Through comprehensive experiments, we validate the robustness of GROOT policies against perceptual variations in simulated and real-world environments. GROOT's performance excels in generalization over background changes, camera viewpoint shifts, and the presence of new object instances, whereas both state-of-the-art end-to-end learning methods and object proposal-based approaches fall short. We also extensively evaluate GROOT policies on real robots, where we demonstrate the efficacy under very wild changes in setup. More videos and model details can be found in the appendix and the project website: https://ut-austin-rpl.github.io/GROOT .

Joydeep Biswas, Sheila Schoepp, Gautham Vasan et al.

Scientific peer review faces mounting strain as submission volumes surge, making it increasingly difficult to sustain review quality, consistency, and timeliness. Recent advances in AI have led the community to consider its use in peer review, yet a key unresolved question is whether AI can generate technically sound reviews at real-world conference scale. Here we report the first large-scale field deployment of AI-assisted peer review: every main-track submission at AAAI-26 received one clearly identified AI review from a state-of-the-art system. The system combined frontier models, tool use, and safeguards in a multi-stage process to generate reviews for all 22,977 full-review papers in less than a day. A large-scale survey of AAAI-26 authors and program committee members showed that participants not only found AI reviews useful, but actually preferred them to human reviews on key dimensions such as technical accuracy and research suggestions. We also introduce a novel benchmark and find that our system substantially outperforms a simple LLM-generated review baseline at detecting a variety of scientific weaknesses. Together, these results show that state-of-the-art AI methods can already make meaningful contributions to scientific peer review at conference scale, opening a path toward the next generation of synergistic human-AI teaming for evaluating research.

Varun Kompella, Thomas J. Walsh, Samuel Barrett et al.

Experience replay (ER) is a crucial component of many deep reinforcement learning (RL) systems. However, uniform sampling from an ER buffer can lead to slow convergence and unstable asymptotic behaviors. This paper introduces Stratified Sampling from Event Tables (SSET), which partitions an ER buffer into Event Tables, each capturing important subsequences of optimal behavior. We prove a theoretical advantage over the traditional monolithic buffer approach and combine SSET with an existing prioritized sampling strategy to further improve learning speed and stability. Empirical results in challenging MiniGrid domains, benchmark RL environments, and a high-fidelity car racing simulator demonstrate the advantages and versatility of SSET over existing ER buffer sampling approaches.

Linji Wang, Zifan Xu, Peter Stone et al.

The high cost of real-world data for robotics Reinforcement Learning (RL) leads to the wide usage of simulators. Despite extensive work on building better dynamics models for simulators to match with the real world, there is another, often-overlooked mismatch between simulations and the real world, namely the distribution of available training tasks. Such a mismatch is further exacerbated by existing curriculum learning techniques, which automatically vary the simulation task distribution without considering its relevance to the real world. Considering these challenges, we posit that curriculum learning for robotics RL needs to be grounded in real-world task distributions. To this end, we propose Grounded Curriculum Learning (GCL), which aligns the simulated task distribution in the curriculum with the real world, as well as explicitly considers what tasks have been given to the robot and how the robot has performed in the past. We validate GCL using the BARN dataset on complex navigation tasks, achieving a 6.8% and 6.5% higher success rate compared to a state-of-the-art CL method and a curriculum designed by human experts, respectively. These results show that GCL can enhance learning efficiency and navigation performance by grounding the simulation task distribution in the real world within an adaptive curriculum.

Zichao Hu, Zifan Xu, Dongsik Chang et al.

The ability to push large objects in a goal-directed manner using onboard egocentric perception is an essential skill for humanoid robots to perform complex tasks such as material handling in warehouses. To robustly manipulate heavy objects to arbitrary goal configurations, the robot must cope with unknown object mass and ground friction, noisy onboard perception, and actuation errors; all in a real-time feedback loop. Existing solutions either rely on privileged object-state information without onboard perception or lack robustness to variations in goal configurations and object physical properties. In this work, we present VOFA, a visual goal-conditioned humanoid loco-manipulation system capable of pushing objects with unknown physical properties to arbitrary goal positions. VOFA consists of a two-level hierarchical architecture with a high-level visuomotor policy and a low-level force-adaptive whole-body controller. The high-level policy processes noisy onboard observations and generates goal-conditioned commands to operate in closed loop across diverse object-goal configurations, while the low-level whole-body controller provides robustness to variations in object physical properties. VOFA is extensively evaluated in both simulation and real-world experiments on the Booster T1 humanoid robot. Our results demonstrate strong performance, achieving over 90% success in simulation and over 80% success in real-world trials. Moreover, VOFA successfully pushes objects weighing up to 17kg, exceeding half of the Booster T1's body weight.

Arrasy Rahman, Jiaxun Cui, Peter Stone

Robustly cooperating with unseen agents and human partners presents significant challenges due to the diverse cooperative conventions these partners may adopt. Existing Ad Hoc Teamwork (AHT) methods address this challenge by training an agent with a population of diverse teammate policies obtained through maximizing specific diversity metrics. However, prior heuristic-based diversity metrics do not always maximize the agent's robustness in all cooperative problems. In this work, we first propose that maximizing an AHT agent's robustness requires it to emulate policies in the minimum coverage set (MCS), the set of best-response policies to any partner policies in the environment. We then introduce the L-BRDiv algorithm that generates a set of teammate policies that, when used for AHT training, encourage agents to emulate policies from the MCS. L-BRDiv works by solving a constrained optimization problem to jointly train teammate policies for AHT training and approximating AHT agent policies that are members of the MCS. We empirically demonstrate that L-BRDiv produces more robust AHT agents than state-of-the-art methods in a broader range of two-player cooperative problems without the need for extensive hyperparameter tuning for its objectives. Our study shows that L-BRDiv outperforms the baseline methods by prioritizing discovering distinct members of the MCS instead of repeatedly finding redundant policies.

Dustin Morrill, Thomas J. Walsh, Daniel Hernandez et al.

Modern reinforcement learning systems produce many high-quality policies throughout the learning process. However, to choose which policy to actually deploy in the real world, they must be tested under an intractable number of environmental conditions. We introduce RPOSST, an algorithm to select a small set of test cases from a larger pool based on a relatively small number of sample evaluations. RPOSST treats the test case selection problem as a two-player game and optimizes a solution with provable $k$-of-$N$ robustness, bounding the error relative to a test that used all the test cases in the pool. Empirical results demonstrate that RPOSST finds a small set of test cases that identify high quality policies in a toy one-shot game, poker datasets, and a high-fidelity racing simulator.

Jiaheng Hu, Zizhao Wang, Peter Stone et al.

Tasks with large state space and sparse rewards present a longstanding challenge to reinforcement learning. In these tasks, an agent needs to explore the state space efficiently until it finds a reward. To deal with this problem, the community has proposed to augment the reward function with intrinsic reward, a bonus signal that encourages the agent to visit interesting states. In this work, we propose a new way of defining interesting states for environments with factored state spaces and complex chained dependencies, where an agent's actions may change the value of one entity that, in order, may affect the value of another entity. Our insight is that, in these environments, interesting states for exploration are states where the agent is uncertain whether (as opposed to how) entities such as the agent or objects have some influence on each other. We present ELDEN, Exploration via Local DepENdencies, a novel intrinsic reward that encourages the discovery of new interactions between entities. ELDEN utilizes a novel scheme -- the partial derivative of the learned dynamics to model the local dependencies between entities accurately and computationally efficiently. The uncertainty of the predicted dependencies is then used as an intrinsic reward to encourage exploration toward new interactions. We evaluate the performance of ELDEN on four different domains with complex dependencies, ranging from 2D grid worlds to 3D robotic tasks. In all domains, ELDEN correctly identifies local dependencies and learns successful policies, significantly outperforming previous state-of-the-art exploration methods.

Siddhant Agarwal, Ishan Durugkar, Peter Stone et al.

Goal-Conditioned Reinforcement Learning (RL) problems often have access to sparse rewards where the agent receives a reward signal only when it has achieved the goal, making policy optimization a difficult problem. Several works augment this sparse reward with a learned dense reward function, but this can lead to sub-optimal policies if the reward is misaligned. Moreover, recent works have demonstrated that effective shaping rewards for a particular problem can depend on the underlying learning algorithm. This paper introduces a novel way to encourage exploration called $f$-Policy Gradients, or $f$-PG. $f$-PG minimizes the f-divergence between the agent's state visitation distribution and the goal, which we show can lead to an optimal policy. We derive gradients for various f-divergences to optimize this objective. Our learning paradigm provides dense learning signals for exploration in sparse reward settings. We further introduce an entropy-regularized policy optimization objective, that we call $state$-MaxEnt RL (or $s$-MaxEnt RL) as a special case of our objective. We show that several metric-based shaping rewards like L2 can be used with $s$-MaxEnt RL, providing a common ground to study such metric-based shaping rewards with efficient exploration. We find that $f$-PG has better performance compared to standard policy gradient methods on a challenging gridworld as well as the Point Maze and FetchReach environments. More information on our website https://agarwalsiddhant10.github.io/projects/fpg.html.

Saman Khamesian, Sri Harini Balaji, Di Yang Shi et al.

Type 1 Diabetes (T1D) management requires continuous adjustment of insulin and lifestyle behaviors to maintain blood glucose within a safe target range. Although automated insulin delivery (AID) systems have improved glycemic outcomes, many patients still fail to achieve recommended clinical targets, warranting new approaches to improve glucose control in patients with T1D. While reinforcement learning (RL) has been utilized as a promising approach, current RL-based methods focus primarily on insulin-only treatment and do not provide behavioral recommendations for glucose control. To address this gap, we propose GUIDE, an RL-based decision-support framework designed to complement AID technologies by providing behavioral recommendations to prevent abnormal glucose events. GUIDE generates structured actions defined by intervention type, magnitude, and timing, including bolus insulin administration and carbohydrate intake events. GUIDE integrates a patient-specific glucose level predictor trained on real-world continuous glucose monitoring data and supports both offline and online RL algorithms within a unified environment. We evaluate both off-policy and on-policy methods across 25 individuals with T1D using standardized glycemic metrics. Among the evaluated approaches, the CQL-BC algorithm demonstrates the highest average time-in-range, reaching 85.49% while maintaining low hypoglycemia exposures. Behavioral similarity analysis further indicates that the learned CQL-BC policy preserves key structural characteristics of patient action patterns, achieving a mean cosine similarity of 0.87 $\pm$ 0.09 across subjects. These findings suggest that conservative offline RL with a structured behavioral action space can provide clinically meaningful and behaviorally plausible decision support for personalized diabetes management.

Carson Stark, Bohkyung Chun, Casey Charleston et al.

This work introduces a robotics platform which embeds a conversational AI agent in an embodied system for natural language understanding and intelligent decision-making for service tasks; integrating task planning and human-like conversation. The agent is derived from a large language model, which has learned from a vast corpus of general knowledge. In addition to generating dialogue, this agent can interface with the physical world by invoking commands on the robot; seamlessly merging communication and behavior. This system is demonstrated in a free-form tour-guide scenario, in an HRI study combining robots with and without conversational AI capabilities. Performance is measured along five dimensions: overall effectiveness, exploration abilities, scrutinization abilities, receptiveness to personification, and adaptability.

Bram Grooten, Patrick MacAlpine, Kaushik Subramanian et al.

Generalization to unseen environments is a significant challenge in the field of robotics and control. In this work, we focus on contextual reinforcement learning, where agents act within environments with varying contexts, such as self-driving cars or quadrupedal robots that need to operate in different terrains or weather conditions than they were trained for. We tackle the critical task of generalizing to out-of-distribution (OOD) settings, without access to explicit context information at test time. Recent work has addressed this problem by training a context encoder and a history adaptation module in separate stages. While promising, this two-phase approach is cumbersome to implement and train. We simplify the methodology and introduce SPARC: single-phase adaptation for robust control. We test SPARC on varying contexts within the high-fidelity racing simulator Gran Turismo 7 and wind-perturbed MuJoCo environments, and find that it achieves reliable and robust OOD generalization.

Zizhao Wang, Chang Shi, Jiaheng Hu et al.

Learning latent actions from action-free video has emerged as a powerful paradigm for scaling up controllable world model learning. Latent actions provide a natural interface for users to iteratively generate and manipulate videos. However, most existing approaches rely on monolithic inverse and forward dynamics models that learn a single latent action to control the entire scene, and therefore struggle in complex environments where multiple entities act simultaneously. This paper introduces Factored Latent Action Model (FLAM), a factored dynamics framework that decomposes the scene into independent factors, each inferring its own latent action and predicting its own next-step factor value. This factorized structure enables more accurate modeling of complex multi-entity dynamics and improves video generation quality in action-free video settings compared to monolithic models. Based on experiments on both simulation and real-world multi-entity datasets, we find that FLAM outperforms prior work in prediction accuracy and representation quality, and facilitates downstream policy learning, demonstrating the benefits of factorized latent action models.

Sveta Paster, Kantwon Rogers, Gordon Briggs et al.

With the projected surge in the elderly population, service robots offer a promising avenue to enhance their well-being in elderly care homes. Such robots will encounter complex scenarios which will require them to perform decisions with ethical consequences. In this report, we propose to leverage the Intelligent Disobedience framework in order to give the robot the ability to perform a deliberation process over decisions with potential ethical implications. We list the issues that this framework can assist with, define it formally in the context of the specific elderly care home scenario, and delineate the requirements for implementing an intelligently disobeying robot. We conclude this report with some critical analysis and suggestions for future work.

Hojoon Lee, Youngdo Lee, Takuma Seno et al.

Scaling up the model size and computation has brought consistent performance improvements in supervised learning. However, this lesson often fails to apply to reinforcement learning (RL) because training the model on non-stationary data easily leads to overfitting and unstable optimization. In response, we introduce SimbaV2, a novel RL architecture designed to stabilize optimization by (i) constraining the growth of weight and feature norm by hyperspherical normalization; and (ii) using a distributional value estimation with reward scaling to maintain stable gradients under varying reward magnitudes. Using the soft actor-critic as a base algorithm, SimbaV2 scales up effectively with larger models and greater compute, achieving state-of-the-art performance on 57 continuous control tasks across 4 domains. The code is available at https://dojeon-ai.github.io/SimbaV2.

Zifan Xu, Ran Gong, Maria Vittoria Minniti et al.

Learning generalizable and robust behavior cloning policies requires large volumes of high-quality robotics data. While human demonstrations (e.g., through teleoperation) serve as the standard source for expert behaviors, acquiring such data at scale in the real world is prohibitively expensive. This paper introduces ExpertGen, a framework that automates expert policy learning in simulation to enable scalable sim-to-real transfer. ExpertGen first initializes a behavior prior using a diffusion policy trained on imperfect demonstrations, which may be synthesized by large language models or provided by humans. Reinforcement learning is then used to steer this prior toward high task success by optimizing the diffusion model's initial noise while keep original policy frozen. By keeping the pretrained diffusion policy frozen, ExpertGen regularizes exploration to remain within safe, human-like behavior manifolds, while also enabling effective learning with only sparse rewards. Empirical evaluations on challenging manipulation benchmarks demonstrate that ExpertGen reliably produces high-quality expert policies with no reward engineering. On industrial assembly tasks, ExpertGen achieves a 90.5% overall success rate, while on long-horizon manipulation tasks it attains 85% overall success, outperforming all baseline methods. The resulting policies exhibit dexterous control and remain robust across diverse initial configurations and failure states. To validate sim-to-real transfer, the learned state-based expert policies are further distilled into visuomotor policies via DAgger and successfully deployed on real robotic hardware.

William Yue, Bo Liu, Peter Stone

Deep generative replay has emerged as a promising approach for continual learning in decision-making tasks. This approach addresses the problem of catastrophic forgetting by leveraging the generation of trajectories from previously encountered tasks to augment the current dataset. However, existing deep generative replay methods for continual learning rely on autoregressive models, which suffer from compounding errors in the generated trajectories. In this paper, we propose a simple, scalable, and non-autoregressive method for continual learning in decision-making tasks using a generative model that generates task samples conditioned on the trajectory timestep. We evaluate our method on Continual World benchmarks and find that our approach achieves state-of-the-art performance on the average success rate metric among continual learning methods. Code is available at https://github.com/WilliamYue37/t-DGR.

Alexander Levine, Peter Stone, Amy Zhang

In real-world control settings, the observation space is often unnecessarily high-dimensional and subject to time-correlated noise. However, the controllable dynamics of the system are often far simpler than the dynamics of the raw observations. It is therefore desirable to learn an encoder to map the observation space to a simpler space of control-relevant variables. In this work, we consider the Ex-BMDP model, first proposed by Efroni et al. (2022), which formalizes control problems where observations can be factorized into an action-dependent latent state which evolves deterministically, and action-independent time-correlated noise. Lamb et al. (2022) proposes the "AC-State" method for learning an encoder to extract a complete action-dependent latent state representation from the observations in such problems. AC-State is a multistep-inverse method, in that it uses the encoding of the the first and last state in a path to predict the first action in the path. However, we identify cases where AC-State will fail to learn a correct latent representation of the agent-controllable factor of the state. We therefore propose a new algorithm, ACDF, which combines multistep-inverse prediction with a latent forward model. ACDF is guaranteed to correctly infer an action-dependent latent state encoder for a large class of Ex-BMDP models. We demonstrate the effectiveness of ACDF on tabular Ex-BMDPs through numerical simulations; as well as high-dimensional environments using neural-network-based encoders. Code is available at https://github.com/midi-lab/acdf.

William Yue, Bo Liu, Peter Stone

In Partially Observable Markov Decision Processes, integrating an agent's history into memory poses a significant challenge for decision-making. Traditional imitation learning, relying on observation-action pairs for expert demonstrations, fails to capture the expert's memory mechanisms used in decision-making. To capture memory processes as demonstrations, we introduce the concept of memory dependency pairs $(p, q)$ indicating that events at time $p$ are recalled for decision-making at time $q$. We introduce AttentionTuner to leverage memory dependency pairs in Transformers and find significant improvements across several tasks compared to standard Transformers when evaluated on Memory Gym and the Long-term Memory Benchmark. Code is available at https://github.com/WilliamYue37/AttentionTuner.