Glen Berseth

Vedant Shah, Johan Obando-Ceron, Vineet Jain et al. · mila

The reasoning performance of large language models (LLMs) can be substantially improved by training them with reinforcement learning (RL). The RL objective for LLM training involves a regularization term, which is the reverse Kullback-Leibler (KL) divergence between the trained policy and the reference policy. Since computing the KL divergence exactly is intractable, various estimators are used in practice to estimate it from on-policy samples. Despite its wide adoption, including in several open-source libraries, there is no systematic study analyzing the numerous ways of incorporating KL estimators in the objective and their effect on the downstream performance of RL-trained models. Recent works show that prevailing practices for incorporating KL regularization do not provide correct gradients for stated objectives, creating a discrepancy between the objective and its implementation. In this paper, we further analyze these practices and study the gradients of several estimators configurations, revealing how design choices shape gradient bias. We substantiate these findings with empirical observations by RL fine-tuning \texttt{Qwen2.5-7B}, \texttt{Llama-3.1-8B-Instruct} and \texttt{Qwen3-4B-Instruct-2507} with different configurations and evaluating their performance on both in- and out-of-distribution tasks. Through our analysis, we observe that, in on-policy settings: (1) estimator configurations with biased gradients can result in training instabilities; and (2) using estimator configurations resulting in unbiased gradients leads to better performance on in-domain as well as out-of-domain tasks. We also investigate the performance resulting from different KL configurations in off-policy settings and observe that KL regularization can help stabilize off-policy RL training resulting from asynchronous setups.

Yandong Ji, Zhongyu Li, Yinan Sun et al.

We address the problem of enabling quadrupedal robots to perform precise shooting skills in the real world using reinforcement learning. Developing algorithms to enable a legged robot to shoot a soccer ball to a given target is a challenging problem that combines robot motion control and planning into one task. To solve this problem, we need to consider the dynamics limitation and motion stability during the control of a dynamic legged robot. Moreover, we need to consider motion planning to shoot the hard-to-model deformable ball rolling on the ground with uncertain friction to a desired location. In this paper, we propose a hierarchical framework that leverages deep reinforcement learning to train (a) a robust motion control policy that can track arbitrary motions and (b) a planning policy to decide the desired kicking motion to shoot a soccer ball to a target. We deploy the proposed framework on an A1 quadrupedal robot and enable it to accurately shoot the ball to random targets in the real world.

Zhongyu Li, Xue Bin Peng, Pieter Abbeel et al.

This work aims to push the limits of agility for bipedal robots by enabling a torque-controlled bipedal robot to perform robust and versatile dynamic jumps in the real world. We present a reinforcement learning framework for training a robot to accomplish a large variety of jumping tasks, such as jumping to different locations and directions. To improve performance on these challenging tasks, we develop a new policy structure that encodes the robot's long-term input/output (I/O) history while also providing direct access to a short-term I/O history. In order to train a versatile jumping policy, we utilize a multi-stage training scheme that includes different training stages for different objectives. After multi-stage training, the policy can be directly transferred to a real bipedal Cassie robot. Training on different tasks and exploring more diverse scenarios lead to highly robust policies that can exploit the diverse set of learned maneuvers to recover from perturbations or poor landings during real-world deployment. Such robustness in the proposed policy enables Cassie to succeed in completing a variety of challenging jump tasks in the real world, such as standing long jumps, jumping onto elevated platforms, and multi-axes jumps.

Zhiyuan Sun, Haochen Shi, Marc-Alexandre Côté et al. · microsoft-research

Large language models (LLMs) have been increasingly applied to tasks in language understanding and interactive decision-making, with their impressive performance largely attributed to the extensive domain knowledge embedded within them. However, the depth and breadth of this knowledge can vary across domains. Many existing approaches assume that LLMs possess a comprehensive understanding of their environment, often overlooking potential gaps in their grasp of actual world dynamics. To address this, we introduce Discover, Verify, and Evolve (DiVE), a framework that discovers world dynamics from a small number of demonstrations, verifies the accuracy of these dynamics, and evolves new, advanced dynamics tailored to the current situation. Through extensive evaluations, we assess the impact of each component on performance and compare the dynamics generated by DiVE to human-annotated dynamics. Our results show that LLMs guided by DiVE make more informed decisions, achieving rewards comparable to human players in the Crafter environment and surpassing methods that require prior task-specific training in the MiniHack environment.

Jensen Gao, Siddharth Reddy, Glen Berseth et al.

We aim to help users communicate their intent to machines using flexible, adaptive interfaces that translate arbitrary user input into desired actions. In this work, we focus on assistive typing applications in which a user cannot operate a keyboard, but can instead supply other inputs, such as webcam images that capture eye gaze or neural activity measured by a brain implant. Standard methods train a model on a fixed dataset of user inputs, then deploy a static interface that does not learn from its mistakes; in part, because extracting an error signal from user behavior can be challenging. We investigate a simple idea that would enable such interfaces to improve over time, with minimal additional effort from the user: online learning from user feedback on the accuracy of the interface's actions. In the typing domain, we leverage backspaces as feedback that the interface did not perform the desired action. We propose an algorithm called x-to-text (X2T) that trains a predictive model of this feedback signal, and uses this model to fine-tune any existing, default interface for translating user input into actions that select words or characters. We evaluate X2T through a small-scale online user study with 12 participants who type sentences by gazing at their desired words, a large-scale observational study on handwriting samples from 60 users, and a pilot study with one participant using an electrocorticography-based brain-computer interface. The results show that X2T learns to outperform a non-adaptive default interface, stimulates user co-adaptation to the interface, personalizes the interface to individual users, and can leverage offline data collected from the default interface to improve its initial performance and accelerate online learning.

Arnav Kumar Jain, Lucas Lehnert, Irina Rish et al.

Animals have a developed ability to explore that aids them in important tasks such as locating food, exploring for shelter, and finding misplaced items. These exploration skills necessarily track where they have been so that they can plan for finding items with relative efficiency. Contemporary exploration algorithms often learn a less efficient exploration strategy because they either condition only on the current state or simply rely on making random open-loop exploratory moves. In this work, we propose $ηψ$-Learning, a method to learn efficient exploratory policies by conditioning on past episodic experience to make the next exploratory move. Specifically, $ηψ$-Learning learns an exploration policy that maximizes the entropy of the state visitation distribution of a single trajectory. Furthermore, we demonstrate how variants of the predecessor representation and successor representations can be combined to predict the state visitation entropy. Our experiments demonstrate the efficacy of $ηψ$-Learning to strategically explore the environment and maximize the state coverage with limited samples.

Jensen Gao, Siddharth Reddy, Glen Berseth et al.

Adaptive interfaces can help users perform sequential decision-making tasks like robotic teleoperation given noisy, high-dimensional command signals (e.g., from a brain-computer interface). Recent advances in human-in-the-loop machine learning enable such systems to improve by interacting with users, but tend to be limited by the amount of data that they can collect from individual users in practice. In this paper, we propose a reinforcement learning algorithm to address this by training an interface to map raw command signals to actions using a combination of offline pre-training and online fine-tuning. To address the challenges posed by noisy command signals and sparse rewards, we develop a novel method for representing and inferring the user's long-term intent for a given trajectory. We primarily evaluate our method's ability to assist users who can only communicate through noisy, high-dimensional input channels through a user study in which 12 participants performed a simulated navigation task by using their eye gaze to modulate a 128-dimensional command signal from their webcam. The results show that our method enables successful goal navigation more often than a baseline directional interface, by learning to denoise user commands signals and provide shared autonomy assistance. We further evaluate on a simulated Sawyer pushing task with eye gaze control, and the Lunar Lander game with simulated user commands, and find that our method improves over baseline interfaces in these domains as well. Extensive ablation experiments with simulated user commands empirically motivate each component of our method.

Brandon Trabucco, Mariano Phielipp, Glen Berseth

The prototypical approach to reinforcement learning involves training policies tailored to a particular agent from scratch for every new morphology. Recent work aims to eliminate the re-training of policies by investigating whether a morphology-agnostic policy, trained on a diverse set of agents with similar task objectives, can be transferred to new agents with unseen morphologies without re-training. This is a challenging problem that required previous approaches to use hand-designed descriptions of the new agent's morphology. Instead of hand-designing this description, we propose a data-driven method that learns a representation of morphology directly from the reinforcement learning objective. Ours is the first reinforcement learning algorithm that can train a policy to generalize to new agent morphologies without requiring a description of the agent's morphology in advance. We evaluate our approach on the standard benchmark for agent-agnostic control, and improve over the current state of the art in zero-shot generalization to new agents. Importantly, our method attains good performance without an explicit description of morphology.

Siddarth Venkatraman, Shivesh Khaitan, Ravi Tej Akella et al.

Offline reinforcement learning (RL) holds promise as a means to learn high-reward policies from a static dataset, without the need for further environment interactions. However, a key challenge in offline RL lies in effectively stitching portions of suboptimal trajectories from the static dataset while avoiding extrapolation errors arising due to a lack of support in the dataset. Existing approaches use conservative methods that are tricky to tune and struggle with multi-modal data (as we show) or rely on noisy Monte Carlo return-to-go samples for reward conditioning. In this work, we propose a novel approach that leverages the expressiveness of latent diffusion to model in-support trajectory sequences as compressed latent skills. This facilitates learning a Q-function while avoiding extrapolation error via batch-constraining. The latent space is also expressive and gracefully copes with multi-modal data. We show that the learned temporally-abstract latent space encodes richer task-specific information for offline RL tasks as compared to raw state-actions. This improves credit assignment and facilitates faster reward propagation during Q-learning. Our method demonstrates state-of-the-art performance on the D4RL benchmarks, particularly excelling in long-horizon, sparse-reward tasks.

Roger Creus Castanyer, Joshua Romoff, Glen Berseth

Exploration bonuses in reinforcement learning guide long-horizon exploration by defining custom intrinsic objectives. Several exploration objectives like count-based bonuses, pseudo-counts, and state-entropy maximization are non-stationary and hence are difficult to optimize for the agent. While this issue is generally known, it is usually omitted and solutions remain under-explored. The key contribution of our work lies in transforming the original non-stationary rewards into stationary rewards through an augmented state representation. For this purpose, we introduce the Stationary Objectives For Exploration (SOFE) framework. SOFE requires identifying sufficient statistics for different exploration bonuses and finding an efficient encoding of these statistics to use as input to a deep network. SOFE is based on proposing state augmentations that expand the state space but hold the promise of simplifying the optimization of the agent's objective. We show that SOFE improves the performance of several exploration objectives, including count-based bonuses, pseudo-counts, and state-entropy maximization. Moreover, SOFE outperforms prior methods that attempt to stabilize the optimization of intrinsic objectives. We demonstrate the efficacy of SOFE in hard-exploration problems, including sparse-reward tasks, pixel-based observations, 3D navigation, and procedurally generated environments.

Faisal Mohamed, Catherine Ji, Benjamin Eysenbach et al. · mila

Effective exploration in reinforcement learning requires not only tracking where an agent has been, but also understanding how the agent perceives and represents the world. To learn powerful representations, an agent should actively explore states that contribute to its knowledge of the environment. Temporal representations can capture the information necessary to solve a wide range of potential tasks while avoiding the computational cost associated with full state reconstruction. In this paper, we propose an exploration method that leverages temporal contrastive representations to guide exploration, prioritizing states with unpredictable future outcomes. We demonstrate that such representations can enable the learning of complex exploratory x in locomotion, manipulation, and embodied-AI tasks, revealing capabilities and behaviors that traditionally require extrinsic rewards. Unlike approaches that rely on explicit distance learning or episodic memory mechanisms (e.g., quasimetric-based methods), our method builds directly on temporal similarities, yielding a simpler yet effective strategy for exploration.

Elham Daneshmand, Shafeef Omar, Glen Berseth et al.

Sim-to-real transfer of locomotion policies often leads to performance degradation due to the inevitable sim-to-real gap. Naively fine-tuning these policies directly on hardware is problematic, as it poses risks of mechanical failure and suffers from high sample inefficiency. In this paper, we address the challenge of safely and efficiently fine-tuning reinforcement learning (RL) policies for dynamic locomotion tasks. Specifically, we focus on fine-tuning policies learned in simulation directly on hardware, while explicitly enforcing safety constraints. In doing so, we introduce SLowRL, a framework that combines Low-Rank Adaptation (LoRA) with training-time safety enforcement via a recovery policy. We evaluate our method both in simulation and on a real Unitree Go2 quadruped robot for jump and trot tasks. Experimental results show that our method achieves a $46.5\%$ reduction in fine-tuning time and near-zero safety violations compared to standard proximal policy optimization (PPO) baselines. Notably, we find that a rank-1 adaptation alone is sufficient to recover pre-trained performance in the real world, while maintaining stable and safe real-world fine-tuning. These results demonstrate the practicality of safe, efficient fine-tuning for dynamic real-world robotic applications.

Raj Ghugare, Santiago Miret, Adriana Hugessen et al.

Reinforcement learning (RL) over text representations can be effective for finding high-value policies that can search over graphs. However, RL requires careful structuring of the search space and algorithm design to be effective in this challenge. Through extensive experiments, we explore how different design choices for text grammar and algorithmic choices for training can affect an RL policy's ability to generate molecules with desired properties. We arrive at a new RL-based molecular design algorithm (ChemRLformer) and perform a thorough analysis using 25 molecule design tasks, including computationally complex protein docking simulations. From this analysis, we discover unique insights in this problem space and show that ChemRLformer achieves state-of-the-art performance while being more straightforward than prior work by demystifying which design choices are actually helpful for text-based molecule design.

Hongyao Tang, Glen Berseth

Deep neural networks provide Reinforcement Learning (RL) powerful function approximators to address large-scale decision-making problems. However, these approximators introduce challenges due to the non-stationary nature of RL training. One source of the challenges in RL is that output predictions can churn, leading to uncontrolled changes after each batch update for states not included in the batch. Although such a churn phenomenon exists in each step of network training, how churn occurs and impacts RL remains under-explored. In this work, we start by characterizing churn in a view of Generalized Policy Iteration with function approximation, and we discover a chain effect of churn that leads to a cycle where the churns in value estimation and policy improvement compound and bias the learning dynamics throughout the iteration. Further, we concretize the study and focus on the learning issues caused by the chain effect in different settings, including greedy action deviation in value-based methods, trust region violation in proximal policy optimization, and dual bias of policy value in actor-critic methods. We then propose a method to reduce the chain effect across different settings, called Churn Approximated ReductIoN (CHAIN), which can be easily plugged into most existing DRL algorithms. Our experiments demonstrate the effectiveness of our method in both reducing churn and improving learning performance across online and offline, value-based and policy-based RL settings, as well as a scaling setting.

Li Gu, Zihuan Jiang, Zhixiang Chi et al.

Graphical user interface (GUI)-based mobile agents automate digital tasks on mobile devices by interpreting natural-language instructions and interacting with the screen. While recent methods apply reinforcement learning (RL) to train vision-language-model(VLM) agents in interactive environments with a primary focus on performance, generalization remains underexplored due to the lack of standardized benchmarks and open-source RL systems. In this work, we formalize the problem as a Contextual Markov Decision Process (CMDP) and introduce \textbf{AndroidWorld-Generalization}, a benchmark with three increasingly challenging regimes for evaluating zero-shot generalization to unseen task instances, templates, and applications. We further propose an RL training system that integrates Group Relative Policy Optimization (GRPO) with a scalable rollout collection system, consisting of containerized infrastructure and asynchronous execution % , and error recovery to support reliable and efficient training. Experiments on AndroidWorld-Generalization show that RL enables a 7B-parameter VLM agent to surpass supervised fine-tuning baselines, yielding a 26.1\% improvement on unseen instances but only limited gains on unseen templates (15.7\%) and apps (8.3\%), underscoring the challenges of generalization. As a preliminary step, we demonstrate that few-shot adaptation at test-time improves performance on unseen apps, motivating future research in this direction. To support reproducibility and fair comparison, we open-source the full RL training system, including the environment, task suite, models, prompt configurations, and the underlying infrastructure \footnote{https://github.com/zihuanjiang/AndroidWorld-Generalization}.

Mohamad H. Danesh, Chenhao Li, Amin Abyaneh et al.

World models promise a paradigm shift in robotics, where an agent learns the underlying physics of its environment once to enable efficient planning and behavior learning. However, current world models are often hardware-locked specialists: a model trained on a Boston Dynamics Spot robot fails catastrophically on a Unitree Go1 due to the mismatch in kinematic and dynamic properties, as the model overfits to specific embodiment constraints rather than capturing the universal locomotion dynamics. Consequently, a slight change in actuator dynamics or limb length necessitates training a new model from scratch. In this work, we take a step towards a framework for training a generalizable Quadrupedal World Model (QWM) that disentangles environmental dynamics from robot morphology. We address the limitations of implicit system identification, where treating static physical properties (like mass or limb length) as latent variables to be inferred from motion history creates an adaptation lag that can compromise zero-shot safety and efficiency. Instead, we explicitly condition the generative dynamics on the robot's engineering specifications. By integrating a physical morphology encoder and a reward normalizer, we enable the model to serve as a neural simulator capable of generalizing across morphologies. This capability unlocks zero-shot control across a range of embodiments. We introduce, for the first time, a world model that enables zero-shot generalization to new morphologies for locomotion. While we carefully study the limitations of our method, QWM operates as a distribution-bounded interpolator within the quadrupedal morphology family rather than a universal physics engine, this work represents a significant step toward morphology-conditioned world models for legged locomotion.

Siddarth Venkatraman, Vineet Jain, Sarthak Mittal et al. · mila

Test-time scaling methods improve the capabilities of large language models (LLMs) by increasing the amount of compute used during inference to make a prediction. Inference-time compute can be scaled in parallel by choosing among multiple independent solutions or sequentially through self-refinement. We propose Recursive Self-Aggregation (RSA), a test-time scaling method inspired by evolutionary methods that combines the benefits of both parallel and sequential scaling. Each step of RSA refines a population of candidate reasoning chains through aggregation of subsets to yield a population of improved solutions, which are then used as the candidate pool for the next iteration. RSA exploits the rich information embedded in the reasoning chains -- not just the final answers -- and enables bootstrapping from partially correct intermediate steps within different chains of thought. Empirically, RSA delivers substantial performance gains with increasing compute budgets across diverse tasks, model families and sizes. Notably, RSA enables Qwen3-4B-Instruct-2507 to achieve competitive performance with larger reasoning models, including DeepSeek-R1 and o3-mini (high), while outperforming purely parallel and sequential scaling strategies across AIME-25, HMMT-25, Reasoning Gym, LiveCodeBench-v6, and SuperGPQA. We further demonstrate that training the model to combine solutions via a novel aggregation-aware reinforcement learning approach yields significant performance gains. Code available at https://github.com/HyperPotatoNeo/RSA.

Zhongyu Li, Xue Bin Peng, Pieter Abbeel et al.

This paper presents a comprehensive study on using deep reinforcement learning (RL) to create dynamic locomotion controllers for bipedal robots. Going beyond focusing on a single locomotion skill, we develop a general control solution that can be used for a range of dynamic bipedal skills, from periodic walking and running to aperiodic jumping and standing. Our RL-based controller incorporates a novel dual-history architecture, utilizing both a long-term and short-term input/output (I/O) history of the robot. This control architecture, when trained through the proposed end-to-end RL approach, consistently outperforms other methods across a diverse range of skills in both simulation and the real world. The study also delves into the adaptivity and robustness introduced by the proposed RL system in developing locomotion controllers. We demonstrate that the proposed architecture can adapt to both time-invariant dynamics shifts and time-variant changes, such as contact events, by effectively using the robot's I/O history. Additionally, we identify task randomization as another key source of robustness, fostering better task generalization and compliance to disturbances. The resulting control policies can be successfully deployed on Cassie, a torque-controlled human-sized bipedal robot. This work pushes the limits of agility for bipedal robots through extensive real-world experiments. We demonstrate a diverse range of locomotion skills, including: robust standing, versatile walking, fast running with a demonstration of a 400-meter dash, and a diverse set of jumping skills, such as standing long jumps and high jumps.

Yizhan Li, Florence Cloutier, Sifan Wu et al.

Generating molecules that satisfy precise numeric constraints over multiple physicochemical properties is critical and challenging. Although large language models (LLMs) are expressive, they struggle with precise multi-objective control and numeric reasoning without external structure and feedback. We introduce \textbf{M olGen}, a fragment-level, retrieval-augmented, two-stage framework for molecule generation under multi-property constraints. Stage I : Prototype generation: a multi-agent reasoner performs retrieval-anchored, fragment-level edits to produce a candidate near the feasible region. Stage II : RL-based fine-grained optimization: a fragment-level optimizer trained with Group Relative Policy Optimization (GRPO) applies one- or multi-hop refinements to explicitly minimize the property errors toward our target while regulating edit complexity and deviation from the prototype. A large, automatically curated dataset with reasoning chains of fragment edits and measured property deltas underpins both stages, enabling deterministic, reproducible supervision and controllable multi-hop reasoning. Unlike prior work, our framework better reasons about molecules by leveraging fragments and supports controllable refinement toward numeric targets. Experiments on generation under two sets of property constraints (QED, LogP, Molecular Weight and HOMO, LUMO) show consistent gains in validity and precise satisfaction of multi-property targets, outperforming strong LLMs and graph-based algorithms.

Homayoun Honari, Roger Creus Castanyer, Michael Przystupa et al.

Distributed training and increasing the gradient update frequency are practical strategies to accelerate learning and improve performance, but both exacerbate a central challenge: \textit{policy lag}, which is the mismatch between the behavior policy generating data and the learning policy being updated. Policy lag can hinder the scaling of on-policy learning algorithms to larger problems. In this paper, we identify the sources of policy lag caused by distributed learning and high update frequency. We use the findings to propose \textit{total Variation-based Advantage aligned Constrained policy Optimization (\methodacronym)} as a practical approach to mitigate policy lag. We empirically validate our method and show that it offers better robustness to policy lag in classic RL tasks and a modern RL for LLM math reasoning task.

Bang Liu, Xinfeng Li, Jiayi Zhang et al. · microsoft-research

The advent of large language models (LLMs) has catalyzed a transformative shift in artificial intelligence, paving the way for advanced intelligent agents capable of sophisticated reasoning, robust perception, and versatile action across diverse domains. As these agents increasingly drive AI research and practical applications, their design, evaluation, and continuous improvement present intricate, multifaceted challenges. This book provides a comprehensive overview, framing intelligent agents within modular, brain-inspired architectures that integrate principles from cognitive science, neuroscience, and computational research. We structure our exploration into four interconnected parts. First, we systematically investigate the modular foundation of intelligent agents, systematically mapping their cognitive, perceptual, and operational modules onto analogous human brain functionalities and elucidating core components such as memory, world modeling, reward processing, goal, and emotion. Second, we discuss self-enhancement and adaptive evolution mechanisms, exploring how agents autonomously refine their capabilities, adapt to dynamic environments, and achieve continual learning through automated optimization paradigms. Third, we examine multi-agent systems, investigating the collective intelligence emerging from agent interactions, cooperation, and societal structures. Finally, we address the critical imperative of building safe and beneficial AI systems, emphasizing intrinsic and extrinsic security threats, ethical alignment, robustness, and practical mitigation strategies necessary for trustworthy real-world deployment. By synthesizing modular AI architectures with insights from different disciplines, this survey identifies key research challenges and opportunities, encouraging innovations that harmonize technological advancement with meaningful societal benefit.

Roger Creus Castanyer, Pablo Samuel Castro, Glen Berseth

AI agent research spans a wide spectrum: from RL agents that learn from scratch to foundation model agents that leverage pre-trained knowledge, yet no unified benchmark enables fair comparison across these approaches. We present Agentick, a benchmark for sequential decision-making agents designed to evaluate RL, LLM, VLM, hybrid, and human agents on common ground and to power research on the fundamental challenges of sequential decision-making. Agentick provides 37 procedurally generated tasks across six capability categories, four difficulty levels, and five observation modalities, all exposed through a single Gymnasium-compatible interface. The benchmark ships with a Coding API, oracle reference policies for all tasks, pre-built SFT datasets, a composable agent harness, and a live leaderboard. An evaluation spanning 27 configurations and over 90,000 episodes reveals that no single approach dominates: GPT-5 mini leads overall at 0.309 oracle-normalized score while PPO dominates planning and multi-agent tasks; the reasoning harness multiplies LLM performance by 3-10x; and ASCII observations consistently outperform natural language. These findings highlight the substantial room for improvement that remains across all agent paradigms. Agentick's capability-decomposed, multi-modal design provides the empirical infrastructure needed to drive progress toward general autonomous agents, both as an evaluation framework and as a training ground for RL post-training of foundation models in truly sequential environments.

Pranav Atreya, Karl Pertsch, Tony Lee et al. · nvidia

Comprehensive, unbiased, and comparable evaluation of modern generalist policies is uniquely challenging: existing approaches for robot benchmarking typically rely on heavy standardization, either by specifying fixed evaluation tasks and environments, or by hosting centralized ''robot challenges'', and do not readily scale to evaluating generalist policies across a broad range of tasks and environments. In this work, we propose RoboArena, a new approach for scalable evaluation of generalist robot policies in the real world. Instead of standardizing evaluations around fixed tasks, environments, or locations, we propose to crowd-source evaluations across a distributed network of evaluators. Importantly, evaluators can freely choose the tasks and environments they evaluate on, enabling easy scaling of diversity, but they are required to perform double-blind evaluations over pairs of policies. Then, by aggregating preference feedback from pairwise comparisons across diverse tasks and environments, we can derive a ranking of policies. We instantiate our approach across a network of evaluators at seven academic institutions using the DROID robot platform. Through more than 600 pairwise real-robot evaluation episodes across seven generalist policies, we demonstrate that our crowd-sourced approach can more accurately rank the performance of existing generalist policies than conventional, centralized evaluation approaches, while being more scalable, resilient, and trustworthy. We open our evaluation network to the community and hope that it can enable more accessible comparisons of generalist robot policies.

Siddarth Venkatraman, Mohsin Hasan, Minsu Kim et al.

Any well-behaved generative model over a variable $\mathbf{x}$ can be expressed as a deterministic transformation of an exogenous ('outsourced') Gaussian noise variable $\mathbf{z}$: $\mathbf{x}=f_θ(\mathbf{z})$. In such a model (\eg, a VAE, GAN, or continuous-time flow-based model), sampling of the target variable $\mathbf{x} \sim p_θ(\mathbf{x})$ is straightforward, but sampling from a posterior distribution of the form $p(\mathbf{x}\mid\mathbf{y}) \propto p_θ(\mathbf{x})r(\mathbf{x},\mathbf{y})$, where $r$ is a constraint function depending on an auxiliary variable $\mathbf{y}$, is generally intractable. We propose to amortize the cost of sampling from such posterior distributions with diffusion models that sample a distribution in the noise space ($\mathbf{z}$). These diffusion samplers are trained by reinforcement learning algorithms to enforce that the transformed samples $f_θ(\mathbf{z})$ are distributed according to the posterior in the data space ($\mathbf{x}$). For many models and constraints, the posterior in noise space is smoother than in data space, making it more suitable for amortized inference. Our method enables conditional sampling under unconditional GAN, (H)VAE, and flow-based priors, comparing favorably with other inference methods. We demonstrate the proposed outsourced diffusion sampling in several experiments with large pretrained prior models: conditional image generation, reinforcement learning with human feedback, and protein structure generation.

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.

Roger Creus Castanyer, Johan Obando-Ceron, Lu Li et al. · mila

Scaling deep reinforcement learning networks is challenging and often results in degraded performance, yet the root causes of this failure mode remain poorly understood. Several recent works have proposed mechanisms to address this, but they are often complex and fail to highlight the causes underlying this difficulty. In this work, we conduct a series of empirical analyses which suggest that the combination of non-stationarity with gradient pathologies, due to suboptimal architectural choices, underlie the challenges of scale. We propose a series of direct interventions that stabilize gradient flow, enabling robust performance across a range of network depths and widths. Our interventions are simple to implement and compatible with well-established algorithms, and result in an effective mechanism that enables strong performance even at large scales. We validate our findings on a variety of agents and suites of environments.

Matthew Riemer, Gopeshh Subbaraj, Glen Berseth et al.

Realtime environments change even as agents perform action inference and learning, thus requiring high interaction frequencies to effectively minimize regret. However, recent advances in machine learning involve larger neural networks with longer inference times, raising questions about their applicability in realtime systems where reaction time is crucial. We present an analysis of lower bounds on regret in realtime reinforcement learning (RL) environments to show that minimizing long-term regret is generally impossible within the typical sequential interaction and learning paradigm, but often becomes possible when sufficient asynchronous compute is available. We propose novel algorithms for staggering asynchronous inference processes to ensure that actions are taken at consistent time intervals, and demonstrate that use of models with high action inference times is only constrained by the environment's effective stochasticity over the inference horizon, and not by action frequency. Our analysis shows that the number of inference processes needed scales linearly with increasing inference times while enabling use of models that are multiple orders of magnitude larger than existing approaches when learning from a realtime simulation of Game Boy games such as Pokémon and Tetris.

Cyrus Neary, Omar G. Younis, Artur Kuramshin et al.

Pre-trained vision-language-action (VLA) models offer a promising foundation for generalist robot policies, but often produce brittle behaviors or unsafe failures when deployed zero-shot in out-of-distribution scenarios. We present Vision-Language-Action Planning & Search (VLAPS) -- a novel framework and accompanying algorithms that embed model-based search into the inference procedure of pre-trained VLA policies to improve their performance on robotic tasks. Specifically, our method biases a modified Monte Carlo Tree Search (MCTS) algorithm -- run using a model of the target environment -- using action priors defined by the VLA policy. By using VLA-derived abstractions and priors in model-based search, VLAPS efficiently explores language-conditioned robotics tasks whose search spaces would otherwise be intractably large. Conversely, by integrating model-based search with the VLA policy's inference procedure, VLAPS yields behaviors that are more performant than those obtained by directly following the VLA policy's action predictions. VLAPS offers a principled framework to: i) control test-time compute in VLA models, ii) leverage a priori knowledge of the robotic environment, and iii) integrate established planning and reinforcement learning techniques into the VLA inference process. Across all experiments, VLAPS significantly outperforms VLA-only baselines on language-specified tasks that would otherwise be intractable for uninformed search algorithms, increasing success rates by as much as 67 percentage points.

Hongyao Tang, Johan Obando-Ceron, Pablo Samuel Castro et al. · mila

Plasticity, or the ability of an agent to adapt to new tasks, environments, or distributions, is crucial for continual learning. In this paper, we study the loss of plasticity in deep continual RL from the lens of churn: network output variability for out-of-batch data induced by mini-batch training. We demonstrate that (1) the loss of plasticity is accompanied by the exacerbation of churn due to the gradual rank decrease of the Neural Tangent Kernel (NTK) matrix; (2) reducing churn helps prevent rank collapse and adjusts the step size of regular RL gradients adaptively. Moreover, we introduce Continual Churn Approximated Reduction (C-CHAIN) and demonstrate it improves learning performance and outperforms baselines in a diverse range of continual learning environments on OpenAI Gym Control, ProcGen, DeepMind Control Suite, and MinAtar benchmarks.

Daniel Lawson, Adriana Hugessen, Charlotte Cloutier et al.

While goal-conditioned behavior cloning (GCBC) methods can perform well on in-distribution training tasks, they do not necessarily generalize zero-shot to tasks that require conditioning on novel state-goal pairs, i.e. combinatorial generalization. In part, this limitation can be attributed to a lack of temporal consistency in the state representation learned by BC; if temporally correlated states are properly encoded to similar latent representations, then the out-of-distribution gap for novel state-goal pairs would be reduced. We formalize this notion by demonstrating how encouraging long-range temporal consistency via successor representations (SR) can facilitate generalization. We then propose a simple yet effective representation learning objective, $\text{BYOL-}γ$ for GCBC, which theoretically approximates the successor representation in the finite MDP case through self-predictive representations, and achieves competitive empirical performance across a suite of challenging tasks requiring combinatorial generalization.

Michael Przystupa, Hongyao Tang, Martin Jagersand et al.

Morphology-aware policy learning is a means of enhancing policy sample efficiency by aggregating data from multiple agents. These types of policies have previously been shown to help generalize over dynamic, kinematic, and limb configuration variations between agent morphologies. Unfortunately, these policies still have sub-optimal zero-shot performance compared to end-to-end finetuning on morphologies at deployment. This limitation has ramifications in practical applications such as robotics because further data collection to perform end-to-end finetuning can be computationally expensive. In this work, we investigate combining morphology-aware pretraining with parameter efficient finetuning (PEFT) techniques to help reduce the learnable parameters necessary to specialize a morphology-aware policy to a target embodiment. We compare directly tuning sub-sets of model weights, input learnable adapters, and prefix tuning techniques for online finetuning. Our analysis reveals that PEFT techniques in conjunction with policy pre-training generally help reduce the number of samples to necessary to improve a policy compared to training models end-to-end from scratch. We further find that tuning as few as less than 1% of total parameters will improve policy performance compared the zero-shot performance of the base pretrained a policy.

Glen Berseth

In the era of deep reinforcement learning, making progress is more complex, as the collected experience must be compressed into a deep model for future exploitation and sampling. Many papers have shown that training a deep learning policy under the changing state and action distribution leads to sub-optimal performance, or even collapse. This naturally leads to the concern that even if the community creates improved exploration algorithms or reward objectives, will those improvements fall on the \textit{deaf ears} of optimization difficulties. This work proposes a new \textit{practical} sub-optimality estimator to determine optimization limitations of deep reinforcement learning algorithms. Through experiments across environments and RL algorithms, it is shown that the difference between the best experience generated is 2-3$\times$ better than the policies' learned performance. This large difference indicates that deep RL methods only exploit half of the good experience they generate.

Roger Creus Castanyer, Faisal Mohamed, Pablo Samuel Castro et al. · mila

Reinforcement learning (RL) algorithms are highly sensitive to reward function specification, which remains a central challenge limiting their broad applicability. We present ARM-FM: Automated Reward Machines via Foundation Models, a framework for automated, compositional reward design in RL that leverages the high-level reasoning capabilities of foundation models (FMs). Reward machines (RMs) -- an automata-based formalism for reward specification -- are used as the mechanism for RL objective specification, and are automatically constructed via the use of FMs. The structured formalism of RMs yields effective task decompositions, while the use of FMs enables objective specifications in natural language. Concretely, we (i) use FMs to automatically generate RMs from natural language specifications; (ii) associate language embeddings with each RM automata-state to enable generalization across tasks; and (iii) provide empirical evidence of ARM-FM's effectiveness in a diverse suite of challenging environments, including evidence of zero-shot generalization.

Alexandre Brown, Glen Berseth

Visual reinforcement learning (RL) is challenging due to the need to extract useful representations from high-dimensional inputs while learning effective control from sparse and noisy rewards. Although large perception models exist, integrating them effectively into RL for visual generalization and improved sample efficiency remains difficult. We propose SegDAC, a Segmentation-Driven Actor-Critic method. SegDAC uses Segment Anything (SAM) for object-centric decomposition and YOLO-World to ground the image segmentation process via text inputs. It includes a novel transformer-based architecture that supports a dynamic number of segments at each time step and effectively learns which segments to focus on using online RL, without using human labels. By evaluating SegDAC over a challenging visual generalization benchmark using Maniskill3, which covers diverse manipulation tasks under strong visual perturbations, we demonstrate that SegDAC achieves significantly better visual generalization, doubling prior performance on the hardest setting and matching or surpassing prior methods in sample efficiency across all evaluated tasks.

Luca Scimeca, Siddarth Venkatraman, Moksh Jain et al. · mila

This paper presents a practical application of Relative Trajectory Balance (RTB), a recently introduced off-policy reinforcement learning (RL) objective that can asymptotically solve Bayesian inverse problems optimally. We extend the original work by using RTB to train conditional diffusion model posteriors from pretrained unconditional priors for challenging linear and non-linear inverse problems in vision, and science. We use the objective alongside techniques such as off-policy backtracking exploration to improve training. Importantly, our results show that existing training-free diffusion posterior methods struggle to perform effective posterior inference in latent space due to inherent biases.

Léa Demeule, Mahtab Sandhu, Glen Berseth

The majority of signal data captured in the real world uses numerous sensors with different resolutions. In practice, however, most deep learning architectures are fixed-resolution; they consider a single resolution at training time and inference time. This is convenient to implement but fails to fully take advantage of the diverse signal data that exists. In contrast, other deep learning architectures are adaptive-resolution; they directly allow various resolutions to be processed at training time and inference time. This benefits robustness and computational efficiency but introduces difficult design constraints that hinder mainstream use. In this work, we address the shortcomings of both fixed-resolution and adaptive-resolution methods by introducing Adaptive Resolution Residual Networks (ARRNs), which inherit the advantages of adaptive-resolution methods and the ease of use of fixed-resolution methods. We construct ARRNs from Laplacian residuals, which serve as generic adaptive-resolution adapters for fixed-resolution layers, and which allow casting high-resolution ARRNs into low-resolution ARRNs at inference time by simply omitting high-resolution Laplacian residuals, thus reducing computational cost on low-resolution signals without compromising performance. We complement this novel component with Laplacian dropout, which regularizes for robustness to a distribution of lower resolutions, and which also regularizes for errors that may be induced by approximate smoothing kernels in Laplacian residuals. We provide a solid grounding for the advantageous properties of ARRNs through a theoretical analysis based on neural operators, and empirically show that ARRNs embrace the challenge posed by diverse resolutions with greater flexibility, robustness, and computational efficiency.

Darshan Patil, Janarthanan Rajendran, Glen Berseth et al. · mila

In the real world, the strong episode resetting mechanisms that are needed to train agents in simulation are unavailable. The \textit{resetting} assumption limits the potential of reinforcement learning in the real world, as providing resets to an agent usually requires the creation of additional handcrafted mechanisms or human interventions. Recent work aims to train agents (\textit{forward}) with learned resets by constructing a second (\textit{backward}) agent that returns the forward agent to the initial state. We find that the termination and timing of the transitions between these two agents are crucial for algorithm success. With this in mind, we create a new algorithm, Reset Free RL with Intelligently Switching Controller (RISC) which intelligently switches between the two agents based on the agent's confidence in achieving its current goal. Our new method achieves state-of-the-art performance on several challenging environments for reset-free RL.

Raj Ghugare, Matthieu Geist, Glen Berseth et al.

Some reinforcement learning (RL) algorithms can stitch pieces of experience to solve a task never seen before during training. This oft-sought property is one of the few ways in which RL methods based on dynamic-programming differ from RL methods based on supervised-learning (SL). Yet, certain RL methods based on off-the-shelf SL algorithms achieve excellent results without an explicit mechanism for stitching; it remains unclear whether those methods forgo this important stitching property. This paper studies this question for the problems of achieving a target goal state and achieving a target return value. Our main result is to show that the stitching property corresponds to a form of combinatorial generalization: after training on a distribution of (state, goal) pairs, one would like to evaluate on (state, goal) pairs not seen together in the training data. Our analysis shows that this sort of generalization is different from i.i.d. generalization. This connection between stitching and generalisation reveals why we should not expect SL-based RL methods to perform stitching, even in the limit of large datasets and models. Based on this analysis, we construct new datasets to explicitly test for this property, revealing that SL-based methods lack this stitching property and hence fail to perform combinatorial generalization. Nonetheless, the connection between stitching and combinatorial generalisation also suggests a simple remedy for improving generalisation in SL: data augmentation. We propose a temporal data augmentation and demonstrate that adding it to SL-based methods enables them to successfully complete tasks not seen together during training. On a high level, this connection illustrates the importance of combinatorial generalization for data efficiency in time-series data beyond tasks beyond RL, like audio, video, or text.

Sean Chen, Jensen Gao, Siddharth Reddy et al.

Building assistive interfaces for controlling robots through arbitrary, high-dimensional, noisy inputs (e.g., webcam images of eye gaze) can be challenging, especially when it involves inferring the user's desired action in the absence of a natural 'default' interface. Reinforcement learning from online user feedback on the system's performance presents a natural solution to this problem, and enables the interface to adapt to individual users. However, this approach tends to require a large amount of human-in-the-loop training data, especially when feedback is sparse. We propose a hierarchical solution that learns efficiently from sparse user feedback: we use offline pre-training to acquire a latent embedding space of useful, high-level robot behaviors, which, in turn, enables the system to focus on using online user feedback to learn a mapping from user inputs to desired high-level behaviors. The key insight is that access to a pre-trained policy enables the system to learn more from sparse rewards than a naïve RL algorithm: using the pre-trained policy, the system can make use of successful task executions to relabel, in hindsight, what the user actually meant to do during unsuccessful executions. We evaluate our method primarily through a user study with 12 participants who perform tasks in three simulated robotic manipulation domains using a webcam and their eye gaze: flipping light switches, opening a shelf door to reach objects inside, and rotating a valve. The results show that our method successfully learns to map 128-dimensional gaze features to 7-dimensional joint torques from sparse rewards in under 10 minutes of online training, and seamlessly helps users who employ different gaze strategies, while adapting to distributional shift in webcam inputs, tasks, and environments.

Glen Berseth, Zhiwei Zhang, Grace Zhang et al.

We develop a new continual meta-learning method to address challenges in sequential multi-task learning. In this setting, the agent's goal is to achieve high reward over any sequence of tasks quickly. Prior meta-reinforcement learning algorithms have demonstrated promising results in accelerating the acquisition of new tasks. However, they require access to all tasks during training. Beyond simply transferring past experience to new tasks, our goal is to devise continual reinforcement learning algorithms that learn to learn, using their experience on previous tasks to learn new tasks more quickly. We introduce a new method, continual meta-policy search (CoMPS), that removes this limitation by meta-training in an incremental fashion, over each task in a sequence, without revisiting prior tasks. CoMPS continuously repeats two subroutines: learning a new task using RL and using the experience from RL to perform completely offline meta-learning to prepare for subsequent task learning. We find that CoMPS outperforms prior continual learning and off-policy meta-reinforcement methods on several sequences of challenging continuous control tasks.

Nicholas Rhinehart, Jenny Wang, Glen Berseth et al.

Humans and animals explore their environment and acquire useful skills even in the absence of clear goals, exhibiting intrinsic motivation. The study of intrinsic motivation in artificial agents is concerned with the following question: what is a good general-purpose objective for an agent? We study this question in dynamic partially-observed environments, and argue that a compact and general learning objective is to minimize the entropy of the agent's state visitation estimated using a latent state-space model. This objective induces an agent to both gather information about its environment, corresponding to reducing uncertainty, and to gain control over its environment, corresponding to reducing the unpredictability of future world states. We instantiate this approach as a deep reinforcement learning agent equipped with a deep variational Bayes filter. We find that our agent learns to discover, represent, and exercise control of dynamic objects in a variety of partially-observed environments sensed with visual observations without extrinsic reward.

Charles Sun, Jędrzej Orbik, Coline Devin et al.

We study how robots can autonomously learn skills that require a combination of navigation and grasping. While reinforcement learning in principle provides for automated robotic skill learning, in practice reinforcement learning in the real world is challenging and often requires extensive instrumentation and supervision. Our aim is to devise a robotic reinforcement learning system for learning navigation and manipulation together, in an autonomous way without human intervention, enabling continual learning under realistic assumptions. Our proposed system, ReLMM, can learn continuously on a real-world platform without any environment instrumentation, without human intervention, and without access to privileged information, such as maps, objects positions, or a global view of the environment. Our method employs a modularized policy with components for manipulation and navigation, where manipulation policy uncertainty drives exploration for the navigation controller, and the manipulation module provides rewards for navigation. We evaluate our method on a room cleanup task, where the robot must navigate to and pick up items scattered on the floor. After a grasp curriculum training phase, ReLMM can learn navigation and grasping together fully automatically, in around 40 hours of autonomous real-world training.

Arnaud Fickinger, Natasha Jaques, Samyak Parajuli et al.

Unsupervised reinforcement learning (RL) studies how to leverage environment statistics to learn useful behaviors without the cost of reward engineering. However, a central challenge in unsupervised RL is to extract behaviors that meaningfully affect the world and cover the range of possible outcomes, without getting distracted by inherently unpredictable, uncontrollable, and stochastic elements in the environment. To this end, we propose an unsupervised RL method designed for high-dimensional, stochastic environments based on an adversarial game between two policies (which we call Explore and Control) controlling a single body and competing over the amount of observation entropy the agent experiences. The Explore agent seeks out states that maximally surprise the Control agent, which in turn aims to minimize surprise, and thereby manipulate the environment to return to familiar and predictable states. The competition between these two policies drives them to seek out increasingly surprising parts of the environment while learning to gain mastery over them. We show formally that the resulting algorithm maximizes coverage of the underlying state in block MDPs with stochastic observations, providing theoretical backing to our hypothesis that this procedure avoids uncontrollable and stochastic distractions. Our experiments further demonstrate that Adversarial Surprise leads to the emergence of complex and meaningful skills, and outperforms state-of-the-art unsupervised reinforcement learning methods in terms of both exploration and zero-shot transfer to downstream tasks.

Soroush Nasiriany, Vitchyr H. Pong, Ashvin Nair et al.

Can we use reinforcement learning to learn general-purpose policies that can perform a wide range of different tasks, resulting in flexible and reusable skills? Contextual policies provide this capability in principle, but the representation of the context determines the degree of generalization and expressivity. Categorical contexts preclude generalization to entirely new tasks. Goal-conditioned policies may enable some generalization, but cannot capture all tasks that might be desired. In this paper, we propose goal distributions as a general and broadly applicable task representation suitable for contextual policies. Goal distributions are general in the sense that they can represent any state-based reward function when equipped with an appropriate distribution class, while the particular choice of distribution class allows us to trade off expressivity and learnability. We develop an off-policy algorithm called distribution-conditioned reinforcement learning (DisCo RL) to efficiently learn these policies. We evaluate DisCo RL on a variety of robot manipulation tasks and find that it significantly outperforms prior methods on tasks that require generalization to new goal distributions.

Zhongyu Li, Xuxin Cheng, Xue Bin Peng et al.

Developing robust walking controllers for bipedal robots is a challenging endeavor. Traditional model-based locomotion controllers require simplifying assumptions and careful modelling; any small errors can result in unstable control. To address these challenges for bipedal locomotion, we present a model-free reinforcement learning framework for training robust locomotion policies in simulation, which can then be transferred to a real bipedal Cassie robot. To facilitate sim-to-real transfer, domain randomization is used to encourage the policies to learn behaviors that are robust across variations in system dynamics. The learned policies enable Cassie to perform a set of diverse and dynamic behaviors, while also being more robust than traditional controllers and prior learning-based methods that use residual control. We demonstrate this on versatile walking behaviors such as tracking a target walking velocity, walking height, and turning yaw.

John D. Co-Reyes, Suvansh Sanjeev, Glen Berseth et al.

Much of the current work on reinforcement learning studies episodic settings, where the agent is reset between trials to an initial state distribution, often with well-shaped reward functions. Non-episodic settings, where the agent must learn through continuous interaction with the world without resets, and where the agent receives only delayed and sparse reward signals, is substantially more difficult, but arguably more realistic considering real-world environments do not present the learner with a convenient "reset mechanism" and easy reward shaping. In this paper, instead of studying algorithmic improvements that can address such non-episodic and sparse reward settings, we instead study the kinds of environment properties that can make learning under such conditions easier. Understanding how properties of the environment impact the performance of reinforcement learning agents can help us to structure our tasks in ways that make learning tractable. We first discuss what we term "environment shaping" -- modifications to the environment that provide an alternative to reward shaping, and may be easier to implement. We then discuss an even simpler property that we refer to as "dynamism," which describes the degree to which the environment changes independent of the agent's actions and can be measured by environment transition entropy. Surprisingly, we find that even this property can substantially alleviate the challenges associated with non-episodic RL in sparse reward settings. We provide an empirical evaluation on a set of new tasks focused on non-episodic learning with sparse rewards. Through this study, we hope to shift the focus of the community towards analyzing how properties of the environment can affect learning and the ultimate type of behavior that is learned via RL.

Brian Yang, Dinesh Jayaraman, Glen Berseth et al.

Existing approaches for visuomotor robotic control typically require characterizing the robot in advance by calibrating the camera or performing system identification. We propose MAVRIC, an approach that works with minimal prior knowledge of the robot's morphology, and requires only a camera view containing the robot and its environment and an unknown control interface. MAVRIC revolves around a mutual information-based method for self-recognition, which discovers visual "control points" on the robot body within a few seconds of exploratory interaction, and these control points in turn are then used for visual servoing. MAVRIC can control robots with imprecise actuation, no proprioceptive feedback, unknown morphologies including novel tools, unknown camera poses, and even unsteady handheld cameras. We demonstrate our method on visually-guided 3D point reaching, trajectory following, and robot-to-robot imitation.

Glen Berseth, Daniel Geng, Coline Devin et al.

Every living organism struggles against disruptive environmental forces to carve out and maintain an orderly niche. We propose that such a struggle to achieve and preserve order might offer a principle for the emergence of useful behaviors in artificial agents. We formalize this idea into an unsupervised reinforcement learning method called surprise minimizing reinforcement learning (SMiRL). SMiRL alternates between learning a density model to evaluate the surprise of a stimulus, and improving the policy to seek more predictable stimuli. The policy seeks out stable and repeatable situations that counteract the environment's prevailing sources of entropy. This might include avoiding other hostile agents, or finding a stable, balanced pose for a bipedal robot in the face of disturbance forces. We demonstrate that our surprise minimizing agents can successfully play Tetris, Doom, control a humanoid to avoid falls, and navigate to escape enemies in a maze without any task-specific reward supervision. We further show that SMiRL can be used together with standard task rewards to accelerate reward-driven learning.

Abdul Rahman Kreidieh, Glen Berseth, Brandon Trabucco et al.

Hierarchies of temporally decoupled policies present a promising approach for enabling structured exploration in complex long-term planning problems. To fully achieve this approach an end-to-end training paradigm is needed. However, training these multi-level policies has had limited success due to challenges arising from interactions between the goal-assigning and goal-achieving levels within a hierarchy. In this article, we consider the policy optimization process as a multi-agent process. This allows us to draw on connections between communication and cooperation in multi-agent RL, and demonstrate the benefits of increased cooperation between sub-policies on the training performance of the overall policy. We introduce a simple yet effective technique for inducing inter-level cooperation by modifying the objective function and subsequent gradients of higher-level policies. Experimental results on a wide variety of simulated robotics and traffic control tasks demonstrate that inducing cooperation results in stronger performing policies and increased sample efficiency on a set of difficult long time horizon tasks. We also find that goal-conditioned policies trained using our method display better transfer to new tasks, highlighting the benefits of our method in learning task-agnostic lower-level behaviors. Videos and code are available at: https://sites.google.com/berkeley.edu/cooperative-hrl.

Ashvin Nair, Shikhar Bahl, Alexander Khazatsky et al.

While reinforcement learning provides an appealing formalism for learning individual skills, a general-purpose robotic system must be able to master an extensive repertoire of behaviors. Instead of learning a large collection of skills individually, can we instead enable a robot to propose and practice its own behaviors automatically, learning about the affordances and behaviors that it can perform in its environment, such that it can then repurpose this knowledge once a new task is commanded by the user? In this paper, we study this question in the context of self-supervised goal-conditioned reinforcement learning. A central challenge in this learning regime is the problem of goal setting: in order to practice useful skills, the robot must be able to autonomously set goals that are feasible but diverse. When the robot's environment and available objects vary, as they do in most open-world settings, the robot must propose to itself only those goals that it can accomplish in its present setting with the objects that are at hand. Previous work only studies self-supervised goal-conditioned RL in a single-environment setting, where goal proposals come from the robot's past experience or a generative model are sufficient. In more diverse settings, this frequently leads to impossible goals and, as we show experimentally, prevents effective learning. We propose a conditional goal-setting model that aims to propose goals that are feasible from the robot's current state. We demonstrate that this enables self-supervised goal-conditioned off-policy learning with raw image observations in the real world, enabling a robot to manipulate a variety of objects and generalize to new objects that were not seen during training.