Pulkit Agrawal

Tao Chen, Megha Tippur, Siyang Wu et al. · deepmind

In-hand object reorientation is necessary for performing many dexterous manipulation tasks, such as tool use in less structured environments that remain beyond the reach of current robots. Prior works built reorientation systems assuming one or many of the following: reorienting only specific objects with simple shapes, limited range of reorientation, slow or quasistatic manipulation, simulation-only results, the need for specialized and costly sensor suites, and other constraints which make the system infeasible for real-world deployment. We present a general object reorientation controller that does not make these assumptions. It uses readings from a single commodity depth camera to dynamically reorient complex and new object shapes by any rotation in real-time, with the median reorientation time being close to seven seconds. The controller is trained using reinforcement learning in simulation and evaluated in the real world on new object shapes not used for training, including the most challenging scenario of reorienting objects held in the air by a downward-facing hand that must counteract gravity during reorientation. Our hardware platform only uses open-source components that cost less than five thousand dollars. Although we demonstrate the ability to overcome assumptions in prior work, there is ample scope for improving absolute performance. For instance, the challenging duck-shaped object not used for training was dropped in 56 percent of the trials. When it was not dropped, our controller reoriented the object within 0.4 radians (23 degrees) 75 percent of the time. Videos are available at: https://taochenshh.github.io/projects/visual-dexterity.

Anurag Ajay, Yilun Du, Abhi Gupta et al. · mit

Recent improvements in conditional generative modeling have made it possible to generate high-quality images from language descriptions alone. We investigate whether these methods can directly address the problem of sequential decision-making. We view decision-making not through the lens of reinforcement learning (RL), but rather through conditional generative modeling. To our surprise, we find that our formulation leads to policies that can outperform existing offline RL approaches across standard benchmarks. By modeling a policy as a return-conditional diffusion model, we illustrate how we may circumvent the need for dynamic programming and subsequently eliminate many of the complexities that come with traditional offline RL. We further demonstrate the advantages of modeling policies as conditional diffusion models by considering two other conditioning variables: constraints and skills. Conditioning on a single constraint or skill during training leads to behaviors at test-time that can satisfy several constraints together or demonstrate a composition of skills. Our results illustrate that conditional generative modeling is a powerful tool for decision-making.

Anurag Ajay, Seungwook Han, Yilun Du et al. · mit

To make effective decisions in novel environments with long-horizon goals, it is crucial to engage in hierarchical reasoning across spatial and temporal scales. This entails planning abstract subgoal sequences, visually reasoning about the underlying plans, and executing actions in accordance with the devised plan through visual-motor control. We propose Compositional Foundation Models for Hierarchical Planning (HiP), a foundation model which leverages multiple expert foundation model trained on language, vision and action data individually jointly together to solve long-horizon tasks. We use a large language model to construct symbolic plans that are grounded in the environment through a large video diffusion model. Generated video plans are then grounded to visual-motor control, through an inverse dynamics model that infers actions from generated videos. To enable effective reasoning within this hierarchy, we enforce consistency between the models via iterative refinement. We illustrate the efficacy and adaptability of our approach in three different long-horizon table-top manipulation tasks.

Dibya Ghosh, Anurag Ajay, Pulkit Agrawal et al. · berkeley

Offline RL algorithms must account for the fact that the dataset they are provided may leave many facets of the environment unknown. The most common way to approach this challenge is to employ pessimistic or conservative methods, which avoid behaviors that are too dissimilar from those in the training dataset. However, relying exclusively on conservatism has drawbacks: performance is sensitive to the exact degree of conservatism, and conservative objectives can recover highly suboptimal policies. In this work, we propose that offline RL methods should instead be adaptive in the presence of uncertainty. We show that acting optimally in offline RL in a Bayesian sense involves solving an implicit POMDP. As a result, optimal policies for offline RL must be adaptive, depending not just on the current state but rather all the transitions seen so far during evaluation.We present a model-free algorithm for approximating this optimal adaptive policy, and demonstrate the efficacy of learning such adaptive policies in offline RL benchmarks.

Gabriel B Margolis, Ge Yang, Kartik Paigwar et al. · deepmind

Agile maneuvers such as sprinting and high-speed turning in the wild are challenging for legged robots. We present an end-to-end learned controller that achieves record agility for the MIT Mini Cheetah, sustaining speeds up to 3.9 m/s. This system runs and turns fast on natural terrains like grass, ice, and gravel and responds robustly to disturbances. Our controller is a neural network trained in simulation via reinforcement learning and transferred to the real world. The two key components are (i) an adaptive curriculum on velocity commands and (ii) an online system identification strategy for sim-to-real transfer leveraged from prior work. Videos of the robot's behaviors are available at: https://agility.csail.mit.edu/

Anthony Simeonov, Yilun Du, Lin Yen-Chen et al. · mit

We present a method for performing tasks involving spatial relations between novel object instances initialized in arbitrary poses directly from point cloud observations. Our framework provides a scalable way for specifying new tasks using only 5-10 demonstrations. Object rearrangement is formalized as the question of finding actions that configure task-relevant parts of the object in a desired alignment. This formalism is implemented in three steps: assigning a consistent local coordinate frame to the task-relevant object parts, determining the location and orientation of this coordinate frame on unseen object instances, and executing an action that brings these frames into the desired alignment. We overcome the key technical challenge of determining task-relevant local coordinate frames from a few demonstrations by developing an optimization method based on Neural Descriptor Fields (NDFs) and a single annotated 3D keypoint. An energy-based learning scheme to model the joint configuration of the objects that satisfies a desired relational task further improves performance. The method is tested on three multi-object rearrangement tasks in simulation and on a real robot. Project website, videos, and code: https://anthonysimeonov.github.io/r-ndf/

Anurag Ajay, Abhishek Gupta, Dibya Ghosh et al. · berkeley

Meta-reinforcement learning algorithms provide a data-driven way to acquire policies that quickly adapt to many tasks with varying rewards or dynamics functions. However, learned meta-policies are often effective only on the exact task distribution on which they were trained and struggle in the presence of distribution shift of test-time rewards or transition dynamics. In this work, we develop a framework for meta-RL algorithms that are able to behave appropriately under test-time distribution shifts in the space of tasks. Our framework centers on an adaptive approach to distributional robustness that trains a population of meta-policies to be robust to varying levels of distribution shift. When evaluated on a potentially shifted test-time distribution of tasks, this allows us to choose the meta-policy with the most appropriate level of robustness, and use it to perform fast adaptation. We formally show how our framework allows for improved regret under distribution shift, and empirically show its efficacy on simulated robotics problems under a wide range of distribution shifts.

Sameer Pai, Tao Chen, Megha Tippur et al. · deepmind

We study the problem of object retrieval in scenarios where visual sensing is absent, object shapes are unknown beforehand and objects can move freely, like grabbing objects out of a drawer. Successful solutions require localizing free objects, identifying specific object instances, and then grasping the identified objects, only using touch feedback. Unlike vision, where cameras can observe the entire scene, touch sensors are local and only observe parts of the scene that are in contact with the manipulator. Moreover, information gathering via touch sensors necessitates applying forces on the touched surface which may disturb the scene itself. Reasoning with touch, therefore, requires careful exploration and integration of information over time -- a challenge we tackle. We present a system capable of using sparse tactile feedback from fingertip touch sensors on a dexterous hand to localize, identify and grasp novel objects without any visual feedback. Videos are available at https://taochenshh.github.io/projects/tactofind.

Anthony Simeonov, Ankit Goyal, Lucas Manuelli et al. · nvidia

We propose a system for rearranging objects in a scene to achieve a desired object-scene placing relationship, such as a book inserted in an open slot of a bookshelf. The pipeline generalizes to novel geometries, poses, and layouts of both scenes and objects, and is trained from demonstrations to operate directly on 3D point clouds. Our system overcomes challenges associated with the existence of many geometrically-similar rearrangement solutions for a given scene. By leveraging an iterative pose de-noising training procedure, we can fit multi-modal demonstration data and produce multi-modal outputs while remaining precise and accurate. We also show the advantages of conditioning on relevant local geometric features while ignoring irrelevant global structure that harms both generalization and precision. We demonstrate our approach on three distinct rearrangement tasks that require handling multi-modality and generalization over object shape and pose in both simulation and the real world. Project website, code, and videos: https://anthonysimeonov.github.io/rpdiff-multi-modal/

Eric Chen, Zhang-Wei Hong, Joni Pajarinen et al.

State-of-the-art reinforcement learning (RL) algorithms typically use random sampling (e.g., $ε$-greedy) for exploration, but this method fails on hard exploration tasks like Montezuma's Revenge. To address the challenge of exploration, prior works incentivize exploration by rewarding the agent when it visits novel states. Such intrinsic rewards (also called exploration bonus or curiosity) often lead to excellent performance on hard exploration tasks. However, on easy exploration tasks, the agent gets distracted by intrinsic rewards and performs unnecessary exploration even when sufficient task (also called extrinsic) reward is available. Consequently, such an overly curious agent performs worse than an agent trained with only task reward. Such inconsistency in performance across tasks prevents the widespread use of intrinsic rewards with RL algorithms. We propose a principled constrained optimization procedure called Extrinsic-Intrinsic Policy Optimization (EIPO) that automatically tunes the importance of the intrinsic reward: it suppresses the intrinsic reward when exploration is unnecessary and increases it when exploration is required. The results is superior exploration that does not require manual tuning in balancing the intrinsic reward against the task reward. Consistent performance gains across sixty-one ATARI games validate our claim. The code is available at https://github.com/Improbable-AI/eipo.

Zhang-Wei Hong, Aviral Kumar, Sathwik Karnik et al.

Offline policy learning is aimed at learning decision-making policies using existing datasets of trajectories without collecting additional data. The primary motivation for using reinforcement learning (RL) instead of supervised learning techniques such as behavior cloning is to find a policy that achieves a higher average return than the trajectories constituting the dataset. However, we empirically find that when a dataset is dominated by suboptimal trajectories, state-of-the-art offline RL algorithms do not substantially improve over the average return of trajectories in the dataset. We argue this is due to an assumption made by current offline RL algorithms of staying close to the trajectories in the dataset. If the dataset primarily consists of sub-optimal trajectories, this assumption forces the policy to mimic the suboptimal actions. We overcome this issue by proposing a sampling strategy that enables the policy to only be constrained to ``good data" rather than all actions in the dataset (i.e., uniform sampling). We present a realization of the sampling strategy and an algorithm that can be used as a plug-and-play module in standard offline RL algorithms. Our evaluation demonstrates significant performance gains in 72 imbalanced datasets, D4RL dataset, and across three different offline RL algorithms. Code is available at https://github.com/Improbable-AI/dw-offline-rl.

Meenal Parakh, Alisha Fong, Anthony Simeonov et al. · deepmind

Large Language Models (LLMs) have been shown to act like planners that can decompose high-level instructions into a sequence of executable instructions. However, current LLM-based planners are only able to operate with a fixed set of skills. We overcome this critical limitation and present a method for using LLM-based planners to query new skills and teach robots these skills in a data and time-efficient manner for rigid object manipulation. Our system can re-use newly acquired skills for future tasks, demonstrating the potential of open world and lifelong learning. We evaluate the proposed framework on multiple tasks in simulation and the real world. Videos are available at: https://sites.google.com/mit.edu/halp-robot-learning.

Zhang-Wei Hong, Pulkit Agrawal, Rémi Tachet des Combes et al.

Most offline reinforcement learning (RL) algorithms return a target policy maximizing a trade-off between (1) the expected performance gain over the behavior policy that collected the dataset, and (2) the risk stemming from the out-of-distribution-ness of the induced state-action occupancy. It follows that the performance of the target policy is strongly related to the performance of the behavior policy and, thus, the trajectory return distribution of the dataset. We show that in mixed datasets consisting of mostly low-return trajectories and minor high-return trajectories, state-of-the-art offline RL algorithms are overly restrained by low-return trajectories and fail to exploit high-performing trajectories to the fullest. To overcome this issue, we show that, in deterministic MDPs with stochastic initial states, the dataset sampling can be re-weighted to induce an artificial dataset whose behavior policy has a higher return. This re-weighted sampling strategy may be combined with any offline RL algorithm. We further analyze that the opportunity for performance improvement over the behavior policy correlates with the positive-sided variance of the returns of the trajectories in the dataset. We empirically show that while CQL, IQL, and TD3+BC achieve only a part of this potential policy improvement, these same algorithms combined with our reweighted sampling strategy fully exploit the dataset. Furthermore, we empirically demonstrate that, despite its theoretical limitation, the approach may still be efficient in stochastic environments. The code is available at https://github.com/Improbable-AI/harness-offline-rl.

Marcel Torne, Max Balsells, Zihan Wang et al. · deepmind

Exploration and reward specification are fundamental and intertwined challenges for reinforcement learning. Solving sequential decision-making tasks requiring expansive exploration requires either careful design of reward functions or the use of novelty-seeking exploration bonuses. Human supervisors can provide effective guidance in the loop to direct the exploration process, but prior methods to leverage this guidance require constant synchronous high-quality human feedback, which is expensive and impractical to obtain. In this work, we present a technique called Human Guided Exploration (HuGE), which uses low-quality feedback from non-expert users that may be sporadic, asynchronous, and noisy. HuGE guides exploration for reinforcement learning not only in simulation but also in the real world, all without meticulous reward specification. The key concept involves bifurcating human feedback and policy learning: human feedback steers exploration, while self-supervised learning from the exploration data yields unbiased policies. This procedure can leverage noisy, asynchronous human feedback to learn policies with no hand-crafted reward design or exploration bonuses. HuGE is able to learn a variety of challenging multi-stage robotic navigation and manipulation tasks in simulation using crowdsourced feedback from non-expert users. Moreover, this paradigm can be scaled to learning directly on real-world robots, using occasional, asynchronous feedback from human supervisors.

Gabriel B Margolis, Pulkit Agrawal

Learned locomotion policies can rapidly adapt to diverse environments similar to those experienced during training but lack a mechanism for fast tuning when they fail in an out-of-distribution test environment. This necessitates a slow and iterative cycle of reward and environment redesign to achieve good performance on a new task. As an alternative, we propose learning a single policy that encodes a structured family of locomotion strategies that solve training tasks in different ways, resulting in Multiplicity of Behavior (MoB). Different strategies generalize differently and can be chosen in real-time for new tasks or environments, bypassing the need for time-consuming retraining. We release a fast, robust open-source MoB locomotion controller, Walk These Ways, that can execute diverse gaits with variable footswing, posture, and speed, unlocking diverse downstream tasks: crouching, hopping, high-speed running, stair traversal, bracing against shoves, rhythmic dance, and more. Video and code release: https://gmargo11.github.io/walk-these-ways/

Jaedong Hwang, Zhang-Wei Hong, Eric Chen et al. · mit

Animals and robots navigate through environments by building and refining maps of space. These maps enable functions including navigation back to home, planning, search and foraging. Here, we use observations from neuroscience, specifically the observed fragmentation of grid cell map in compartmentalized spaces, to propose and apply the concept of Fragmentation-and-Recall (FARMap) in the mapping of large spaces. Agents solve the mapping problem by building local maps via a surprisal-based clustering of space, which they use to set subgoals for spatial exploration. Agents build and use a local map to predict their observations; high surprisal leads to a "fragmentation event" that truncates the local map. At these events, the recent local map is placed into long-term memory (LTM) and a different local map is initialized. If observations at a fracture point match observations in one of the stored local maps, that map is recalled (and thus reused) from LTM. The fragmentation points induce a natural online clustering of the larger space, forming a set of intrinsic potential subgoals that are stored in LTM as a topological graph. Agents choose their next subgoal from the set of near and far potential subgoals from within the current local map or LTM, respectively. Thus, local maps guide exploration locally, while LTM promotes global exploration. We demonstrate that FARMap replicates the fragmentation points observed in animal studies. We evaluate FARMap on complex procedurally-generated spatial environments and realistic simulations to demonstrate that this mapping strategy much more rapidly covers the environment (number of agent steps and wall clock time) and is more efficient in active memory usage, without loss of performance. https://jd730.github.io/projects/FARMap/

Yandong Ji, Gabriel B. Margolis, Pulkit Agrawal

DribbleBot (Dexterous Ball Manipulation with a Legged Robot) is a legged robotic system that can dribble a soccer ball under the same real-world conditions as humans (i.e., in-the-wild). We adopt the paradigm of training policies in simulation using reinforcement learning and transferring them into the real world. We overcome critical challenges of accounting for variable ball motion dynamics on different terrains and perceiving the ball using body-mounted cameras under the constraints of onboard computing. Our results provide evidence that current quadruped platforms are well-suited for studying dynamic whole-body control problems involving simultaneous locomotion and manipulation directly from sensory observations.

Haokuan Luo, Albert Yue, Zhang-Wei Hong et al.

We present a strong baseline that surpasses the performance of previously published methods on the Habitat Challenge task of navigating to a target object in indoor environments. Our method is motivated from primary failure modes of prior state-of-the-art: poor exploration, inaccurate object identification, and agent getting trapped due to imprecise map construction. We make three contributions to mitigate these issues: (i) First, we show that existing map-based methods fail to effectively use semantic clues for exploration. We present a semantic-agnostic exploration strategy (called Stubborn) without any learning that surprisingly outperforms prior work. (ii) We propose a strategy for integrating temporal information to improve object identification. (iii) Lastly, due to inaccurate depth observation the agent often gets trapped in small regions. We develop a multi-scale collision map for obstacle identification that mitigates this issue.

Aviv Netanyahu, Tianmin Shu, Joshua Tenenbaum et al.

In this work, we consider one-shot imitation learning for object rearrangement tasks, where an AI agent needs to watch a single expert demonstration and learn to perform the same task in different environments. To achieve a strong generalization, the AI agent must infer the spatial goal specification for the task. However, there can be multiple goal specifications that fit the given demonstration. To address this, we propose a reward learning approach, Graph-based Equivalence Mappings (GEM), that can discover spatial goal representations that are aligned with the intended goal specification, enabling successful generalization in unseen environments. Specifically, GEM represents a spatial goal specification by a reward function conditioned on i) a graph indicating important spatial relationships between objects and ii) state equivalence mappings for each edge in the graph indicating invariant properties of the corresponding relationship. GEM combines inverse reinforcement learning and active reward learning to efficiently improve the reward function by utilizing the graph structure and domain randomization enabled by the equivalence mappings. We conducted experiments with simulated oracles and with human subjects. The results show that GEM can drastically improve the generalizability of the learned goal representations over strong baselines.

Ligong Han, Seungwook Han, Shivchander Sudalairaj et al.

Transformations based on domain expertise (expert transformations), such as random-resized-crop and color-jitter, have proven critical to the success of contrastive learning techniques such as SimCLR. Recently, several attempts have been made to replace such domain-specific, human-designed transformations with generated views that are learned. However for imagery data, so far none of these view-generation methods has been able to outperform expert transformations. In this work, we tackle a different question: instead of replacing expert transformations with generated views, can we constructively assimilate generated views with expert transformations? We answer this question in the affirmative and propose a view generation method and a simple, effective assimilation method that together improve the state-of-the-art by up to ~3.6% on three different datasets. Importantly, we conduct a detailed empirical study that systematically analyzes a range of view generation and assimilation methods and provides a holistic picture of the efficacy of learned views in contrastive representation learning.

Changling Li, Zhang-Wei Hong, Pulkit Agrawal et al.

Experience replay serves as a key component in the success of online reinforcement learning (RL). Prioritized experience replay (PER) reweights experiences by the temporal difference (TD) error empirically enhancing the performance. However, few works have explored the motivation of using TD error. In this work, we provide an alternative perspective on TD-error-based reweighting. We show the connections between the experience prioritization and occupancy optimization. By using a regularized RL objective with $f-$divergence regularizer and employing its dual form, we show that an optimal solution to the objective is obtained by shifting the distribution of off-policy data in the replay buffer towards the on-policy optimal distribution using TD-error-based occupancy ratios. Our derivation results in a new pipeline of TD error prioritization. We specifically explore the KL divergence as the regularizer and obtain a new form of prioritization scheme, the regularized optimal experience replay (ROER). We evaluate the proposed prioritization scheme with the Soft Actor-Critic (SAC) algorithm in continuous control MuJoCo and DM Control benchmark tasks where our proposed scheme outperforms baselines in 6 out of 11 tasks while the results of the rest match with or do not deviate far from the baselines. Further, using pretraining, ROER achieves noticeable improvement on difficult Antmaze environment where baselines fail, showing applicability to offline-to-online fine-tuning. Code is available at \url{https://github.com/XavierChanglingLi/Regularized-Optimal-Experience-Replay}.

Yanwei Wang, Ching-Yun Ko, Pulkit Agrawal

One powerful paradigm in visual navigation is to predict actions from observations directly. Training such an end-to-end system allows representations useful for downstream tasks to emerge automatically. However, the lack of inductive bias makes this system data inefficient. We hypothesize a sufficient representation of the current view and the goal view for a navigation policy can be learned by predicting the location and size of a crop of the current view that corresponds to the goal. We further show that training such random crop prediction in a self-supervised fashion purely on synthetic noise images transfers well to natural home images. The learned representation can then be bootstrapped to learn a navigation policy efficiently with little interaction data. The code is available at https://yanweiw.github.io/noise2ptz

Ge Yang, Anurag Ajay, Pulkit Agrawal

Value approximation using deep neural networks is at the heart of off-policy deep reinforcement learning, and is often the primary module that provides learning signals to the rest of the algorithm. While multi-layer perceptron networks are universal function approximators, recent works in neural kernel regression suggest the presence of a spectral bias, where fitting high-frequency components of the value function requires exponentially more gradient update steps than the low-frequency ones. In this work, we re-examine off-policy reinforcement learning through the lens of kernel regression and propose to overcome such bias via a composite neural tangent kernel. With just a single line-change, our approach, the Fourier feature networks (FFN) produce state-of-the-art performance on challenging continuous control domains with only a fraction of the compute. Faster convergence and better off-policy stability also make it possible to remove the target network without suffering catastrophic divergences, which further reduces TD}(0)'s estimation bias on a few tasks.

Andi Peng, Aviv Netanyahu, Mark Ho et al.

Policies often fail due to distribution shift -- changes in the state and reward that occur when a policy is deployed in new environments. Data augmentation can increase robustness by making the model invariant to task-irrelevant changes in the agent's observation. However, designers don't know which concepts are irrelevant a priori, especially when different end users have different preferences about how the task is performed. We propose an interactive framework to leverage feedback directly from the user to identify personalized task-irrelevant concepts. Our key idea is to generate counterfactual demonstrations that allow users to quickly identify possible task-relevant and irrelevant concepts. The knowledge of task-irrelevant concepts is then used to perform data augmentation and thus obtain a policy adapted to personalized user objectives. We present experiments validating our framework on discrete and continuous control tasks with real human users. Our method (1) enables users to better understand agent failure, (2) reduces the number of demonstrations required for fine-tuning, and (3) aligns the agent to individual user task preferences.

Martin Peticco, Gabriella Ulloa, John Marangola et al.

Development of dexterous manipulation hardware has primarily focused on hands and grippers. However, these end-effectors are often paired with bulky and highly stiff wrists that limit performance in human environments. More designs have adopted backdrivable actuation, but are still difficult to model and control due to coupled kinematics or high mechanical inertia from heavy links. We present DexWrist, a robotic wrist that advances manipulation in highly constrained environments and enables dynamic, contact-rich tasks. We achieve this by combining quasi-direct drive actuation with a decoupled parallel kinematic mechanism in a compact design. It delivers 3.75 +/- 0.05 Nm rated torque, 0.33 +/- 0.06 Nm backdrive torque, 10.15 +/- 1.34 Hz torque bandwidth, +/- 40 degrees ROM in both DOFs, and a one-to-one motor-to-DOF mapping in a 0.97 kg package. In practice, these properties increase workspace in cluttered environments and stabilize contact without the need for finely tuned admittance control. We evaluate DexWrist as a drop-in wrist upgrade in simulation and on two robot arms performing representative constrained and contact-rich tasks. In learned policy evaluations, DexWrist achieved 50-76% relative improvements in success rate, and reduced autonomous task completion times by 3-5x. More details about DexWrist can be found at https://dexwrist.csail.mit.edu.

Idan Shenfeld, Zhang-Wei Hong, Aviv Tamar et al.

Learning from rewards (i.e., reinforcement learning or RL) and learning to imitate a teacher (i.e., teacher-student learning) are two established approaches for solving sequential decision-making problems. To combine the benefits of these different forms of learning, it is common to train a policy to maximize a combination of reinforcement and teacher-student learning objectives. However, without a principled method to balance these objectives, prior work used heuristics and problem-specific hyperparameter searches to balance the two objectives. We present a $\textit{principled}$ approach, along with an approximate implementation for $\textit{dynamically}$ and $\textit{automatically}$ balancing when to follow the teacher and when to use rewards. The main idea is to adjust the importance of teacher supervision by comparing the agent's performance to the counterfactual scenario of the agent learning without teacher supervision and only from rewards. If using teacher supervision improves performance, the importance of teacher supervision is increased and otherwise it is decreased. Our method, $\textit{Teacher Guided Reinforcement Learning}$ (TGRL), outperforms strong baselines across diverse domains without hyper-parameter tuning.

Aviv Netanyahu, Abhishek Gupta, Max Simchowitz et al.

Machine learning systems, especially with overparameterized deep neural networks, can generalize to novel test instances drawn from the same distribution as the training data. However, they fare poorly when evaluated on out-of-support test points. In this work, we tackle the problem of developing machine learning systems that retain the power of overparameterized function approximators while enabling extrapolation to out-of-support test points when possible. This is accomplished by noting that under certain conditions, a "transductive" reparameterization can convert an out-of-support extrapolation problem into a problem of within-support combinatorial generalization. We propose a simple strategy based on bilinear embeddings to enable this type of combinatorial generalization, thereby addressing the out-of-support extrapolation problem under certain conditions. We instantiate a simple, practical algorithm applicable to various supervised learning and imitation learning tasks.

Allen Z. Ren, Justin Lidard, Lars L. Ankile et al.

We introduce Diffusion Policy Policy Optimization, DPPO, an algorithmic framework including best practices for fine-tuning diffusion-based policies (e.g. Diffusion Policy) in continuous control and robot learning tasks using the policy gradient (PG) method from reinforcement learning (RL). PG methods are ubiquitous in training RL policies with other policy parameterizations; nevertheless, they had been conjectured to be less efficient for diffusion-based policies. Surprisingly, we show that DPPO achieves the strongest overall performance and efficiency for fine-tuning in common benchmarks compared to other RL methods for diffusion-based policies and also compared to PG fine-tuning of other policy parameterizations. Through experimental investigation, we find that DPPO takes advantage of unique synergies between RL fine-tuning and the diffusion parameterization, leading to structured and on-manifold exploration, stable training, and strong policy robustness. We further demonstrate the strengths of DPPO in a range of realistic settings, including simulated robotic tasks with pixel observations, and via zero-shot deployment of simulation-trained policies on robot hardware in a long-horizon, multi-stage manipulation task. Website with code: diffusion-ppo.github.io

Andi Peng, Mycal Tucker, Eoin Kenny et al. · mit

Neural networks often learn task-specific latent representations that fail to generalize to novel settings or tasks. Conversely, humans learn discrete representations (i.e., concepts or words) at a variety of abstraction levels (e.g., "bird" vs. "sparrow") and deploy the appropriate abstraction based on task. Inspired by this, we train neural models to generate a spectrum of discrete representations, and control the complexity of the representations (roughly, how many bits are allocated for encoding inputs) by tuning the entropy of the distribution over representations. In finetuning experiments, using only a small number of labeled examples for a new task, we show that (1) tuning the representation to a task-appropriate complexity level supports the highest finetuning performance, and (2) in a human-participant study, users were able to identify the appropriate complexity level for a downstream task using visualizations of discrete representations. Our results indicate a promising direction for rapid model finetuning by leveraging human insight.

Zhang-Wei Hong, Ge Yang, Pulkit Agrawal

The dominant framework for off-policy multi-goal reinforcement learning involves estimating goal conditioned Q-value function. When learning to achieve multiple goals, data efficiency is intimately connected with the generalization of the Q-function to new goals. The de-facto paradigm is to approximate Q(s, a, g) using monolithic neural networks. To improve the generalization of the Q-function, we propose a bilinear decomposition that represents the Q-value via a low-rank approximation in the form of a dot product between two vector fields. The first vector field, f(s, a), captures the environment's local dynamics at the state s; whereas the second component, φ(s, g), captures the global relationship between the current state and the goal. We show that our bilinear decomposition scheme substantially improves data efficiency, and has superior transfer to out-of-distribution goals compared to prior methods. Empirical evidence is provided on the simulated Fetch robot task-suite and dexterous manipulation with a Shadow hand.

Lars Ankile, Anthony Simeonov, Idan Shenfeld et al.

Recent advances in Behavior Cloning (BC) have made it easy to teach robots new tasks. However, we find that the ease of teaching comes at the cost of unreliable performance that saturates with increasing data for tasks requiring precision. The performance saturation can be attributed to two critical factors: (a) distribution shift resulting from the use of offline data and (b) the lack of closed-loop corrective control caused by action chucking (predicting a set of future actions executed open-loop) critical for BC performance. Our key insight is that by predicting action chunks, BC policies function more like trajectory "planners" than closed-loop controllers necessary for reliable execution. To address these challenges, we devise a simple yet effective method, ResiP (Residual for Precise Manipulation), that overcomes the reliability problem while retaining BC's ease of teaching and long-horizon capabilities. ResiP augments a frozen, chunked BC model with a fully closed-loop residual policy trained with reinforcement learning (RL) that addresses distribution shifts and introduces closed-loop corrections over open-loop execution of action chunks predicted by the BC trajectory planner. Videos, code, and data: https://residual-assembly.github.io.

Jatin Kishnani, Mayank Goel, Amit Singh et al.

We present the first end-to-end demonstration of fine-tuning and serving Google's Gemma 4 31B model on TPU hardware, providing an empirical comparison of TPU and GPU platforms for large language model adaptation. Using LoRA on a Google TPU v5p-8 for training and TPU v6e-8 (Trillium) for inference, we document the full set of code-level adaptations required to port a GPU-native training recipe, built on PyTorch, HuggingFace TRL, and FSDP, to the JAX + Tunix/Qwix stack. These adaptations span mesh configuration, LoRA module naming conventions, sharding annotation corrections, gradient checkpointing, data pipeline restructuring, and a custom Orbax-to-safetensors checkpoint merging procedure. For inference, we detail the vLLM-TPU Docker setup necessary to serve Gemma 4 on v6e-8 and characterize the resulting latency and throughput profile. Compared with a 2xH100 GPU baseline under identical hyperparameters, TPU training completes 1.61x faster at 2.12x lower cost. Inference throughput is within 3% across platforms, while TPU achieves 2x lower time-to-first-token (235 ms vs. 475 ms). Together, the TPU configuration is 1.82x cheaper for a representative train-plus-service workload. Our work removes a critical gap in the open tooling ecosystem and provides practitioners with a reproducible, production-ready recipe for Gemma 4 deployment on TPU infrastructure.

Richard Li, Carlos Esteves, Ameesh Makadia et al.

We present a system for accurately predicting stable orientations for diverse rigid objects. We propose to overcome the critical issue of modelling multimodality in the space of rotations by using a conditional generative model to accurately classify contact surfaces. Our system is capable of operating from noisy and partially-observed pointcloud observations captured by real world depth cameras. Our method substantially outperforms the current state-of-the-art systems on a simulated stacking task requiring highly accurate rotations, and demonstrates strong sim2real zero-shot transfer results across a variety of unseen objects on a real world reorientation task. Project website: \url{https://richardrl.github.io/stable-reorientation/}

Jaedong Hwang, Zhang-Wei Hong, Eric Chen et al.

Deep reinforcement learning methods exhibit impressive performance on a range of tasks but still struggle on hard exploration tasks in large environments with sparse rewards. To address this, intrinsic rewards can be generated using forward model prediction errors that decrease as the environment becomes known, and incentivize an agent to explore novel states. While prediction-based intrinsic rewards can help agents solve hard exploration tasks, they can suffer from catastrophic forgetting and actually increase at visited states. We first examine the conditions and causes of catastrophic forgetting in grid world environments. We then propose a new method FARCuriosity, inspired by how humans and animals learn. The method depends on fragmentation and recall: an agent fragments an environment based on surprisal, and uses different local curiosity modules (prediction-based intrinsic reward functions) for each fragment so that modules are not trained on the entire environment. At each fragmentation event, the agent stores the current module in long-term memory (LTM) and either initializes a new module or recalls a previously stored module based on its match with the current state. With fragmentation and recall, FARCuriosity achieves less forgetting and better overall performance in games with varied and heterogeneous environments in the Atari benchmark suite of tasks. Thus, this work highlights the problem of catastrophic forgetting in prediction-based curiosity methods and proposes a solution.

Andreea Bobu, Andi Peng, Pulkit Agrawal et al.

To act in the world, robots rely on a representation of salient task aspects: for example, to carry a coffee mug, a robot may consider movement efficiency or mug orientation in its behavior. However, if we want robots to act for and with people, their representations must not be just functional but also reflective of what humans care about, i.e. they must be aligned. We observe that current learning approaches suffer from representation misalignment, where the robot's learned representation does not capture the human's representation. We suggest that because humans are the ultimate evaluator of robot performance, we must explicitly focus our efforts on aligning learned representations with humans, in addition to learning the downstream task. We advocate that current representation learning approaches in robotics should be studied from the perspective of how well they accomplish the objective of representation alignment. We mathematically define the problem, identify its key desiderata, and situate current methods within this formalism. We conclude by suggesting future directions for exploring open challenges.

Gabriel B. Margolis, Xiang Fu, Yandong Ji et al.

Knowledge of terrain's physical properties inferred from color images can aid in making efficient robotic locomotion plans. However, unlike image classification, it is unintuitive for humans to label image patches with physical properties. Without labeled data, building a vision system that takes as input the observed terrain and predicts physical properties remains challenging. We present a method that overcomes this challenge by self-supervised labeling of images captured by robots during real-world traversal with physical property estimators trained in simulation. To ensure accurate labeling, we introduce Active Sensing Motor Policies (ASMP), which are trained to explore locomotion behaviors that increase the accuracy of estimating physical parameters. For instance, the quadruped robot learns to swipe its foot against the ground to estimate the friction coefficient accurately. We show that the visual system trained with a small amount of real-world traversal data accurately predicts physical parameters. The trained system is robust and works even with overhead images captured by a drone despite being trained on data collected by cameras attached to a quadruped robot walking on the ground.

Tao Chen, Eric Cousineau, Naveen Kuppuswamy et al.

Recent studies have made significant progress in addressing dexterous manipulation problems, particularly in in-hand object reorientation. However, there are few existing works that explore the potential utilization of developed dexterous manipulation controllers for downstream tasks. In this study, we focus on constrained dexterous manipulation for food peeling. Food peeling presents various constraints on the reorientation controller, such as the requirement for the hand to securely hold the object after reorientation for peeling. We propose a simple system for learning a reorientation controller that facilitates the subsequent peeling task. Videos are available at: https://taochenshh.github.io/projects/veg-peeling.

Amit Singh, Vedant Nipane, Pulkit Agrawal et al.

Large language models (LLMs) demonstrate strong code generation abilities in general-purpose programming languages but remain limited in specialized domains such as low-level embedded systems programming. This domain involves hardware register manipulation, vendor-specific SDKs, real-time operating system APIs, and hardware abstraction layers that are underrepresented in standard pretraining corpora. We introduce H2LooP Spark Preview, a continual pretraining (CPT) pipeline that adapts the OLMo-3-7B-a fully open language model to the embedded systems domain using BF16 LoRA with rank-stabilized scaling on 8 NVIDIA H100 GPUs. Our training corpus is constructed from repository-datasheet pairs covering 100B tokens of raw embedded systems data across 117 manufacturers, processed using the hierarchical datasheet-to-code mapping approach proposed in SpecMap (Nipane et al., 2026). The resulting curated dataset split contains 23.5B tokens across 13 embedded domains. Continual pretraining with high-rank LoRA (r=512) yields substantial gains, reducing in-domain perplexity by 70.4% and held-out repository perplexity by 66.1%. On generative code completion benchmarks spanning 13 embedded domains, our 7B model outperforms Claude Opus 4.6 and Qwen3-Coder-30B on 8 categories in token accuracy, showing that targeted continual pretraining enables smaller open-weight models to rival frontier systems on specialized technical tasks. We release the production training checkpoint on Huggingface as an open-source artifact.

Max Balsells, Marcel Torne, Zihan Wang et al.

Ideally, we would place a robot in a real-world environment and leave it there improving on its own by gathering more experience autonomously. However, algorithms for autonomous robotic learning have been challenging to realize in the real world. While this has often been attributed to the challenge of sample complexity, even sample-efficient techniques are hampered by two major challenges - the difficulty of providing well "shaped" rewards, and the difficulty of continual reset-free training. In this work, we describe a system for real-world reinforcement learning that enables agents to show continual improvement by training directly in the real world without requiring painstaking effort to hand-design reward functions or reset mechanisms. Our system leverages occasional non-expert human-in-the-loop feedback from remote users to learn informative distance functions to guide exploration while leveraging a simple self-supervised learning algorithm for goal-directed policy learning. We show that in the absence of resets, it is particularly important to account for the current "reachability" of the exploration policy when deciding which regions of the space to explore. Based on this insight, we instantiate a practical learning system - GEAR, which enables robots to simply be placed in real-world environments and left to train autonomously without interruption. The system streams robot experience to a web interface only requiring occasional asynchronous feedback from remote, crowdsourced, non-expert humans in the form of binary comparative feedback. We evaluate this system on a suite of robotic tasks in simulation and demonstrate its effectiveness at learning behaviors both in simulation and the real world. Project website https://guided-exploration-autonomous-rl.github.io/GEAR/.

Max Simchowitz, Anurag Ajay, Pulkit Agrawal et al.

This paper studies the prediction of a target $\mathbf{z}$ from a pair of random variables $(\mathbf{x},\mathbf{y})$, where the ground-truth predictor is additive $\mathbb{E}[\mathbf{z} \mid \mathbf{x},\mathbf{y}] = f_\star(\mathbf{x}) +g_{\star}(\mathbf{y})$. We study the performance of empirical risk minimization (ERM) over functions $f+g$, $f \in F$ and $g \in G$, fit on a given training distribution, but evaluated on a test distribution which exhibits covariate shift. We show that, when the class $F$ is "simpler" than $G$ (measured, e.g., in terms of its metric entropy), our predictor is more resilient to heterogeneous covariate shifts} in which the shift in $\mathbf{x}$ is much greater than that in $\mathbf{y}$. Our analysis proceeds by demonstrating that ERM behaves qualitatively similarly to orthogonal machine learning: the rate at which ERM recovers the $f$-component of the predictor has only a lower-order dependence on the complexity of the class $G$, adjusted for partial non-indentifiability introduced by the additive structure. These results rely on a novel Hölder style inequality for the Dudley integral which may be of independent interest. Moreover, we corroborate our theoretical findings with experiments demonstrating improved resilience to shifts in "simpler" features across numerous domains.

Idan Shenfeld, Mehul Damani, Jonas Hübotter et al.

Continual learning, enabling models to acquire new skills and knowledge without degrading existing capabilities, remains a fundamental challenge for foundation models. While on-policy reinforcement learning can reduce forgetting, it requires explicit reward functions that are often unavailable. Learning from expert demonstrations, the primary alternative, is dominated by supervised fine-tuning (SFT), which is inherently off-policy. We introduce Self-Distillation Fine-Tuning (SDFT), a simple method that enables on-policy learning directly from demonstrations. SDFT leverages in-context learning by using a demonstration-conditioned model as its own teacher, generating on-policy training signals that preserve prior capabilities while acquiring new skills. Across skill learning and knowledge acquisition tasks, SDFT consistently outperforms SFT, achieving higher new-task accuracy while substantially reducing catastrophic forgetting. In sequential learning experiments, SDFT enables a single model to accumulate multiple skills over time without performance regression, establishing on-policy distillation as a practical path to continual learning from demonstrations.

Srinath Mahankali, Zhang-Wei Hong, Ayush Sekhari et al.

We introduce Random Latent Exploration (RLE), a simple yet effective exploration strategy in reinforcement learning (RL). On average, RLE outperforms noise-based methods, which perturb the agent's actions, and bonus-based exploration, which rewards the agent for attempting novel behaviors. The core idea of RLE is to encourage the agent to explore different parts of the environment by pursuing randomly sampled goals in a latent space. RLE is as simple as noise-based methods, as it avoids complex bonus calculations but retains the deep exploration benefits of bonus-based methods. Our experiments show that RLE improves performance on average in both discrete (e.g., Atari) and continuous control tasks (e.g., Isaac Gym), enhancing exploration while remaining a simple and general plug-in for existing RL algorithms. Project website and code: https://srinathm1359.github.io/random-latent-exploration

Younghyo Park, Gabriel B. Margolis, Pulkit Agrawal

Many roboticists dream of presenting a robot with a task in the evening and returning the next morning to find the robot capable of solving the task. What is preventing us from achieving this? Sim-to-real reinforcement learning (RL) has achieved impressive performance on challenging robotics tasks, but requires substantial human effort to set up the task in a way that is amenable to RL. It's our position that algorithmic improvements in policy optimization and other ideas should be guided towards resolving the primary bottleneck of shaping the training environment, i.e., designing observations, actions, rewards and simulation dynamics. Most practitioners don't tune the RL algorithm, but other environment parameters to obtain a desirable controller. We posit that scaling RL to diverse robotic tasks will only be achieved if the community focuses on automating environment shaping procedures.

Ryan Bahlous-Boldi, Isha Puri, Idan Shenfeld et al.

Language models must now generalize out of the box to novel environments and work inside inference-scaling search procedures, such as AlphaEvolve, that select rollouts with a variety of task-specific reward functions. Unfortunately, the standard paradigm of LLM post-training optimizes a pre-specified scalar reward, often leading current LLMs to produce low-entropy response distributions and thus to struggle at displaying the diversity that inference-time search will require. We propose Vector Policy Optimization (VPO), an RL algorithm that explicitly trains policies to anticipate diverse downstream reward functions and to produce diverse solutions. VPO exploits that rewards are often vector-valued in practice, like per-test-case correctness in code generation or, say, multiple different user personas or reward models. VPO is essentially a drop-in replacement for the GRPO advantage estimator, but it trains the LLM to output a set of solutions where individual solutions specialize to different trade-offs in the vector reward space. Across four tasks, VPO matches or beats the strongest scalar RL baselines on test-time search (e.g. pass@k and best@k), with the gap widening as the search budget grows. For evolutionary search, VPO models unlock problems that GRPO models cannot solve at all. As test-time search becomes more standardized, optimizing for diversity may need to become the default post-training objective.

Tom Gunter, Zirui Wang, Chong Wang et al.

We present foundation language models developed to power Apple Intelligence features, including a ~3 billion parameter model designed to run efficiently on devices and a large server-based language model designed for Private Cloud Compute. These models are designed to perform a wide range of tasks efficiently, accurately, and responsibly. This report describes the model architecture, the data used to train the model, the training process, how the models are optimized for inference, and the evaluation results. We highlight our focus on Responsible AI and how the principles are applied throughout the model development.

Zhang-Wei Hong, Idan Shenfeld, Tsun-Hsuan Wang et al.

Large language models (LLMs) hold great potential for many natural language applications but risk generating incorrect or toxic content. To probe when an LLM generates unwanted content, the current paradigm is to recruit a \textit{red team} of human testers to design input prompts (i.e., test cases) that elicit undesirable responses from LLMs. However, relying solely on human testers is expensive and time-consuming. Recent works automate red teaming by training a separate red team LLM with reinforcement learning (RL) to generate test cases that maximize the chance of eliciting undesirable responses from the target LLM. However, current RL methods are only able to generate a small number of effective test cases resulting in a low coverage of the span of prompts that elicit undesirable responses from the target LLM. To overcome this limitation, we draw a connection between the problem of increasing the coverage of generated test cases and the well-studied approach of curiosity-driven exploration that optimizes for novelty. Our method of curiosity-driven red teaming (CRT) achieves greater coverage of test cases while mantaining or increasing their effectiveness compared to existing methods. Our method, CRT successfully provokes toxic responses from LLaMA2 model that has been heavily fine-tuned using human preferences to avoid toxic outputs. Code is available at \url{https://github.com/Improbable-AI/curiosity_redteam}

Vedant Nipane, Pulkit Agrawal, Amit Singh

Establishing precise traceability between embedded systems datasheets and their corresponding code implementations remains a fundamental challenge in systems engineering, particularly for low-level software where manual mapping between specification documents and large code repositories is infeasible. Existing Traceability Link Recovery approaches primarily rely on lexical similarity and information retrieval techniques, which struggle to capture the semantic, structural, and symbol level relationships prevalent in embedded systems software. We present a hierarchical datasheet-to-code mapping methodology that employs large language models for semantic analysis while explicitly structuring the traceability process across multiple abstraction levels. Rather than performing direct specification-to-code matching, the proposed approach progressively narrows the search space through repository-level structure inference, file-level relevance estimation, and fine-grained symbollevel alignment. The method extends beyond function-centric mapping by explicitly covering macros, structs, constants, configuration parameters, and register definitions commonly found in systems-level C/C++ codebases. We evaluate the approach on multiple open-source embedded systems repositories using manually curated datasheet-to-code ground truth. Experimental results show substantial improvements over traditional information-retrieval-based baselines, achieving up to 73.3% file mapping accuracy. We significantly reduce computational overhead, lowering total LLM token consumption by 84% and end-to-end runtime by approximately 80%. This methodology supports automated analysis of large embedded software systems and enables downstream applications such as training data generation for systems-aware machine learning models, standards compliance verification, and large-scale specification coverage analysis.

Adam Zweiger, Jyothish Pari, Han Guo et al. · mit

Large language models (LLMs) are powerful but static; they lack mechanisms to adapt their weights in response to new tasks, knowledge, or examples. We introduce Self-Adapting LLMs (SEAL), a framework that enables LLMs to self-adapt by generating their own finetuning data and update directives. Given a new input, the model produces a self-edit-a generation that may restructure the information in different ways, specify optimization hyperparameters, or invoke tools for data augmentation and gradient-based updates. Through supervised finetuning (SFT), these self-edits result in persistent weight updates, enabling lasting adaptation. To train the model to produce effective self-edits, we use a reinforcement learning loop with the downstream performance of the updated model as the reward signal. Unlike prior approaches that rely on separate adaptation modules or auxiliary networks, SEAL directly uses the model's own generation to control its adaptation process. Experiments on knowledge incorporation and few-shot generalization show that SEAL is a promising step toward language models capable of self-directed adaptation. Our website and code is available at https://jyopari.github.io/posts/seal.

Martin Peticco, Pulkit Agrawal

Traditional robotic hand metrics focus on static properties such as workspace, manipulability, and grasp stability. However, these metrics do not directly measure dexterity under the standard definition in robotic manipulation: the ability to continuously change an object's pose within the hand while maintaining contact from an initial grasp. We introduce Kinematic Rolling Manipulation Ability (KaRMA), a kinematic-only metric for fine manipulation that quantifies reachable in-hand translation and reorientation of a spherical test object within a two-finger precision pinch through feasible rolling motions. KaRMA enforces joint limits, collision constraints, rolling contact, and antipodal force feasibility, then investigates reachable in-hand object poses via breadth-first search over translation and rotation primitives. KaRMA reports three scores: translational coverage (KaRMA-T), rotational coverage (KaRMA-R), and sensitivity to the initial grasp (KaRMA-S). We evaluate KaRMA on 16 widely used robotic hands and compare against static baselines, showing that KaRMA separates hands that rank identically under static proxies, reveals translation-rotation tradeoffs invisible to existing baselines, and is qualitatively consistent with selected published task benchmarks where Jacobian-based metrics can be misleading.

Antonia Bronars, Younghyo Park, Pulkit Agrawal

Position controllers have become the dominant interface for executing learned manipulation policies. Yet a critical design decision remains understudied: how should we choose controller gains for policy learning? The conventional wisdom is to select gains based on desired task compliance or stiffness. However, this logic breaks down when controllers are paired with state-conditioned policies: effective stiffness emerges from the interplay between learned reactions and control dynamics, not from gains alone. We argue that gain selection should instead be guided by learnability: how amenable different gain settings are to the learning algorithm in use. In this work, we systematically investigate how position controller gains affect three core components of modern robot learning pipelines: behavior cloning, reinforcement learning from scratch, and sim-to-real transfer. Through extensive experiments across multiple tasks and robot embodiments, we find that: (1) behavior cloning benefits from compliant and overdamped gain regimes, (2) reinforcement learning can succeed across all gain regimes given compatible hyperparameter tuning, and (3) sim-to-real transfer is harmed by stiff and overdamped gain regimes. These findings reveal that optimal gain selection depends not on the desired task behavior, but on the learning paradigm employed. Project website: https://younghyopark.me/tune-to-learn