Dorsa Sadigh

Megha Srivastava, Erdem Biyik, Suvir Mirchandani et al. · stanford

Recent works on shared autonomy and assistive-AI technologies, such as assistive robot teleoperation, seek to model and help human users with limited ability in a fixed task. However, these approaches often fail to account for humans' ability to adapt and eventually learn how to execute a control task themselves. Furthermore, in applications where it may be desirable for a human to intervene, these methods may inhibit their ability to learn how to succeed with full self-control. In this paper, we focus on the problem of assistive teaching of motor control tasks such as parking a car or landing an aircraft. Despite their ubiquitous role in humans' daily activities and occupations, motor tasks are rarely taught in a uniform way due to their high complexity and variance. We propose an AI-assisted teaching algorithm that leverages skill discovery methods from reinforcement learning (RL) to (i) break down any motor control task into teachable skills, (ii) construct novel drill sequences, and (iii) individualize curricula to students with different capabilities. Through an extensive mix of synthetic and user studies on two motor control tasks -- parking a car with a joystick and writing characters from the Balinese alphabet -- we show that assisted teaching with skills improves student performance by around 40% compared to practicing full trajectories without skills, and practicing with individualized drills can result in up to 25% further improvement. Our source code is available at https://github.com/Stanford-ILIAD/teaching

Stephen Casper, Xander Davies, Claudia Shi et al. · berkeley, eth-zurich

Reinforcement learning from human feedback (RLHF) is a technique for training AI systems to align with human goals. RLHF has emerged as the central method used to finetune state-of-the-art large language models (LLMs). Despite this popularity, there has been relatively little public work systematizing its flaws. In this paper, we (1) survey open problems and fundamental limitations of RLHF and related methods; (2) overview techniques to understand, improve, and complement RLHF in practice; and (3) propose auditing and disclosure standards to improve societal oversight of RLHF systems. Our work emphasizes the limitations of RLHF and highlights the importance of a multi-faceted approach to the development of safer AI systems.

Siddharth Karamcheti, Suraj Nair, Annie S. Chen et al. · stanford

Recent work in visual representation learning for robotics demonstrates the viability of learning from large video datasets of humans performing everyday tasks. Leveraging methods such as masked autoencoding and contrastive learning, these representations exhibit strong transfer to policy learning for visuomotor control. But, robot learning encompasses a diverse set of problems beyond control including grasp affordance prediction, language-conditioned imitation learning, and intent scoring for human-robot collaboration, amongst others. First, we demonstrate that existing representations yield inconsistent results across these tasks: masked autoencoding approaches pick up on low-level spatial features at the cost of high-level semantics, while contrastive learning approaches capture the opposite. We then introduce Voltron, a framework for language-driven representation learning from human videos and associated captions. Voltron trades off language-conditioned visual reconstruction to learn low-level visual patterns, and visually-grounded language generation to encode high-level semantics. We also construct a new evaluation suite spanning five distinct robot learning problems $\unicode{x2013}$ a unified platform for holistically evaluating visual representations for robotics. Through comprehensive, controlled experiments across all five problems, we find that Voltron's language-driven representations outperform the prior state-of-the-art, especially on targeted problems requiring higher-level features.

Yuchen Cui, Siddharth Karamcheti, Raj Palleti et al. · stanford

Systems for language-guided human-robot interaction must satisfy two key desiderata for broad adoption: adaptivity and learning efficiency. Unfortunately, existing instruction-following agents cannot adapt, lacking the ability to incorporate online natural language supervision, and even if they could, require hundreds of demonstrations to learn even simple policies. In this work, we address these problems by presenting Language-Informed Latent Actions with Corrections (LILAC), a framework for incorporating and adapting to natural language corrections - "to the right," or "no, towards the book" - online, during execution. We explore rich manipulation domains within a shared autonomy paradigm. Instead of discrete turn-taking between a human and robot, LILAC splits agency between the human and robot: language is an input to a learned model that produces a meaningful, low-dimensional control space that the human can use to guide the robot. Each real-time correction refines the human's control space, enabling precise, extended behaviors - with the added benefit of requiring only a handful of demonstrations to learn. We evaluate our approach via a user study where users work with a Franka Emika Panda manipulator to complete complex manipulation tasks. Compared to existing learned baselines covering both open-loop instruction following and single-turn shared autonomy, we show that our corrections-aware approach obtains higher task completion rates, and is subjectively preferred by users because of its reliability, precision, and ease of use.

Suvir Mirchandani, Fei Xia, Pete Florence et al. · stanford

We observe that pre-trained large language models (LLMs) are capable of autoregressively completing complex token sequences -- from arbitrary ones procedurally generated by probabilistic context-free grammars (PCFG), to more rich spatial patterns found in the Abstraction and Reasoning Corpus (ARC), a general AI benchmark, prompted in the style of ASCII art. Surprisingly, pattern completion proficiency can be partially retained even when the sequences are expressed using tokens randomly sampled from the vocabulary. These results suggest that without any additional training, LLMs can serve as general sequence modelers, driven by in-context learning. In this work, we investigate how these zero-shot capabilities may be applied to problems in robotics -- from extrapolating sequences of numbers that represent states over time to complete simple motions, to least-to-most prompting of reward-conditioned trajectories that can discover and represent closed-loop policies (e.g., a stabilizing controller for CartPole). While difficult to deploy today for real systems due to latency, context size limitations, and compute costs, the approach of using LLMs to drive low-level control may provide an exciting glimpse into how the patterns among words could be transferred to actions.

Maximilian Du, Suraj Nair, Dorsa Sadigh et al. · stanford

Enabling robots to learn novel visuomotor skills in a data-efficient manner remains an unsolved problem with myriad challenges. A popular paradigm for tackling this problem is through leveraging large unlabeled datasets that have many behaviors in them and then adapting a policy to a specific task using a small amount of task-specific human supervision (i.e. interventions or demonstrations). However, how best to leverage the narrow task-specific supervision and balance it with offline data remains an open question. Our key insight in this work is that task-specific data not only provides new data for an agent to train on but can also inform the type of prior data the agent should use for learning. Concretely, we propose a simple approach that uses a small amount of downstream expert data to selectively query relevant behaviors from an offline, unlabeled dataset (including many sub-optimal behaviors). The agent is then jointly trained on the expert and queried data. We observe that our method learns to query only the relevant transitions to the task, filtering out sub-optimal or task-irrelevant data. By doing so, it is able to learn more effectively from the mix of task-specific and offline data compared to naively mixing the data or only using the task-specific data. Furthermore, we find that our simple querying approach outperforms more complex goal-conditioned methods by 20% across simulated and real robotic manipulation tasks from images. See https://sites.google.com/view/behaviorretrieval for videos and code.

Hengyuan Hu, Suvir Mirchandani, Dorsa Sadigh · stanford

Despite the considerable potential of reinforcement learning (RL), robotic control tasks predominantly rely on imitation learning (IL) due to its better sample efficiency. However, it is costly to collect comprehensive expert demonstrations that enable IL to generalize to all possible scenarios, and any distribution shift would require recollecting data for finetuning. Therefore, RL is appealing if it can build upon IL as an efficient autonomous self-improvement procedure. We propose imitation bootstrapped reinforcement learning (IBRL), a novel framework for sample-efficient RL with demonstrations that first trains an IL policy on the provided demonstrations and then uses it to propose alternative actions for both online exploration and bootstrapping target values. Compared to prior works that oversample the demonstrations or regularize RL with an additional imitation loss, IBRL is able to utilize high quality actions from IL policies since the beginning of training, which greatly accelerates exploration and training efficiency. We evaluate IBRL on 6 simulation and 3 real-world tasks spanning various difficulty levels. IBRL significantly outperforms prior methods and the improvement is particularly more prominent in harder tasks.

Joey Hejna, Chethan Bhateja, Yichen Jiang et al.

Increasingly large imitation learning datasets are being collected with the goal of training foundation models for robotics. However, despite the fact that data selection has been of utmost importance in vision and natural language processing, little work in robotics has questioned what data such models should actually be trained on. In this work we investigate how to weigh different subsets or ``domains'' of robotics datasets for robot foundation model pre-training. Concrete, we use distributionally robust optimization (DRO) to maximize worst-case performance across all possible downstream domains. Our method, Re-Mix, addresses the wide range of challenges that arise when applying DRO to robotics datasets including variability in action spaces and dynamics across different datasets. Re-Mix employs early stopping, action normalization, and discretization to counteract these issues. Through extensive experimentation on the largest open-source robot manipulation dataset, the Open X-Embodiment dataset, we demonstrate that data curation can have an outsized impact on downstream performance. Specifically, domain weights learned by Re-Mix outperform uniform weights by 38\% on average and outperform human-selected weights by 32\% on datasets used to train existing generalist robot policies, specifically the RT-X models.

Kanishk Gandhi, Siddharth Karamcheti, Madeline Liao et al. · stanford

Imitation learning from human-provided demonstrations is a strong approach for learning policies for robot manipulation. While the ideal dataset for imitation learning is homogenous and low-variance -- reflecting a single, optimal method for performing a task -- natural human behavior has a great deal of heterogeneity, with several optimal ways to demonstrate a task. This multimodality is inconsequential to human users, with task variations manifesting as subconscious choices; for example, reaching down, then across to grasp an object, versus reaching across, then down. Yet, this mismatch presents a problem for interactive imitation learning, where sequences of users improve on a policy by iteratively collecting new, possibly conflicting demonstrations. To combat this problem of demonstrator incompatibility, this work designs an approach for 1) measuring the compatibility of a new demonstration given a base policy, and 2) actively eliciting more compatible demonstrations from new users. Across two simulation tasks requiring long-horizon, dexterous manipulation and a real-world "food plating" task with a Franka Emika Panda arm, we show that we can both identify incompatible demonstrations via post-hoc filtering, and apply our compatibility measure to actively elicit compatible demonstrations from new users, leading to improved task success rates across simulated and real environments.

Minae Kwon, Hengyuan Hu, Vivek Myers et al. · stanford

Consider a robot tasked with tidying a desk with a meticulously constructed Lego sports car. A human may recognize that it is not appropriate to disassemble the sports car and put it away as part of the "tidying." How can a robot reach that conclusion? Although large language models (LLMs) have recently been used to enable commonsense reasoning, grounding this reasoning in the real world has been challenging. To reason in the real world, robots must go beyond passively querying LLMs and actively gather information from the environment that is required to make the right decision. For instance, after detecting that there is an occluded car, the robot may need to actively perceive the car to know whether it is an advanced model car made out of Legos or a toy car built by a toddler. We propose an approach that leverages an LLM and vision language model (VLM) to help a robot actively perceive its environment to perform grounded commonsense reasoning. To evaluate our framework at scale, we release the MessySurfaces dataset which contains images of 70 real-world surfaces that need to be cleaned. We additionally illustrate our approach with a robot on 2 carefully designed surfaces. We find an average 12.9% improvement on the MessySurfaces benchmark and an average 15% improvement on the robot experiments over baselines that do not use active perception. The dataset, code, and videos of our approach can be found at https://minaek.github.io/grounded_commonsense_reasoning.

Minyoung Hwang, Joey Hejna, Dorsa Sadigh et al. · cmu

While success in many robotics tasks can be determined by only observing the final state and how it differs from the initial state - e.g., if an apple is picked up - many tasks require observing the full motion of the robot to correctly determine success. For example, brushing hair requires repeated strokes that correspond to the contours and type of hair. Prior works often use off-the-shelf vision-language models (VLMs) as success detectors; however, when success depends on the full trajectory, VLMs struggle to make correct judgments for two reasons. First, modern VLMs are trained only on single frames, and cannot capture changes over a full trajectory. Second, even if we provide state-of-the-art VLMs with an aggregate input of multiple frames, they still fail to detect success due to a lack of robot data. Our key idea is to fine-tune VLMs using abstract representations that are able to capture trajectory-level information such as the path the robot takes by overlaying keypoint trajectories on the final image. We propose motion instruction fine-tuning (MotIF), a method that fine-tunes VLMs using the aforementioned abstract representations to semantically ground the robot's behavior in the environment. To benchmark and fine-tune VLMs for robotic motion understanding, we introduce the MotIF-1K dataset containing 653 human and 369 robot demonstrations across 13 task categories. MotIF assesses the success of robot motion given the image observation of the trajectory, task instruction, and motion description. Our model significantly outperforms state-of-the-art VLMs by at least twice in precision and 56.1% in recall, generalizing across unseen motions, tasks, and environments. Finally, we demonstrate practical applications of MotIF in refining and terminating robot planning, and ranking trajectories on how they align with task and motion descriptions. Project page: https://motif-1k.github.io

Wenhao Yu, Nimrod Gileadi, Chuyuan Fu et al.

Large language models (LLMs) have demonstrated exciting progress in acquiring diverse new capabilities through in-context learning, ranging from logical reasoning to code-writing. Robotics researchers have also explored using LLMs to advance the capabilities of robotic control. However, since low-level robot actions are hardware-dependent and underrepresented in LLM training corpora, existing efforts in applying LLMs to robotics have largely treated LLMs as semantic planners or relied on human-engineered control primitives to interface with the robot. On the other hand, reward functions are shown to be flexible representations that can be optimized for control policies to achieve diverse tasks, while their semantic richness makes them suitable to be specified by LLMs. In this work, we introduce a new paradigm that harnesses this realization by utilizing LLMs to define reward parameters that can be optimized and accomplish variety of robotic tasks. Using reward as the intermediate interface generated by LLMs, we can effectively bridge the gap between high-level language instructions or corrections to low-level robot actions. Meanwhile, combining this with a real-time optimizer, MuJoCo MPC, empowers an interactive behavior creation experience where users can immediately observe the results and provide feedback to the system. To systematically evaluate the performance of our proposed method, we designed a total of 17 tasks for a simulated quadruped robot and a dexterous manipulator robot. We demonstrate that our proposed method reliably tackles 90% of the designed tasks, while a baseline using primitive skills as the interface with Code-as-policies achieves 50% of the tasks. We further validated our method on a real robot arm where complex manipulation skills such as non-prehensile pushing emerge through our interactive system.

Allen Z. Ren, Anushri Dixit, Alexandra Bodrova et al.

Large language models (LLMs) exhibit a wide range of promising capabilities -- from step-by-step planning to commonsense reasoning -- that may provide utility for robots, but remain prone to confidently hallucinated predictions. In this work, we present KnowNo, which is a framework for measuring and aligning the uncertainty of LLM-based planners such that they know when they don't know and ask for help when needed. KnowNo builds on the theory of conformal prediction to provide statistical guarantees on task completion while minimizing human help in complex multi-step planning settings. Experiments across a variety of simulated and real robot setups that involve tasks with different modes of ambiguity (e.g., from spatial to numeric uncertainties, from human preferences to Winograd schemas) show that KnowNo performs favorably over modern baselines (which may involve ensembles or extensive prompt tuning) in terms of improving efficiency and autonomy, while providing formal assurances. KnowNo can be used with LLMs out of the box without model-finetuning, and suggests a promising lightweight approach to modeling uncertainty that can complement and scale with the growing capabilities of foundation models. Website: https://robot-help.github.io

Minae Kwon, Sang Michael Xie, Kalesha Bullard et al.

Reward design in reinforcement learning (RL) is challenging since specifying human notions of desired behavior may be difficult via reward functions or require many expert demonstrations. Can we instead cheaply design rewards using a natural language interface? This paper explores how to simplify reward design by prompting a large language model (LLM) such as GPT-3 as a proxy reward function, where the user provides a textual prompt containing a few examples (few-shot) or a description (zero-shot) of the desired behavior. Our approach leverages this proxy reward function in an RL framework. Specifically, users specify a prompt once at the beginning of training. During training, the LLM evaluates an RL agent's behavior against the desired behavior described by the prompt and outputs a corresponding reward signal. The RL agent then uses this reward to update its behavior. We evaluate whether our approach can train agents aligned with user objectives in the Ultimatum Game, matrix games, and the DealOrNoDeal negotiation task. In all three tasks, we show that RL agents trained with our framework are well-aligned with the user's objectives and outperform RL agents trained with reward functions learned via supervised learning

Dorsa Sadigh, Eric S. Kim, Samuel Coogan et al.

We propose to synthesize a control policy for a Markov decision process (MDP) such that the resulting traces of the MDP satisfy a linear temporal logic (LTL) property. We construct a product MDP that incorporates a deterministic Rabin automaton generated from the desired LTL property. The reward function of the product MDP is defined from the acceptance condition of the Rabin automaton. This construction allows us to apply techniques from learning theory to the problem of synthesis for LTL specifications even when the transition probabilities are not known a priori. We prove that our method is guaranteed to find a controller that satisfies the LTL property with probability one if such a policy exists, and we suggest empirically with a case study in traffic control that our method produces reasonable control strategies even when the LTL property cannot be satisfied with probability one.

Jensen Gao, Bidipta Sarkar, Fei Xia et al.

Recent advances in vision-language models (VLMs) have led to improved performance on tasks such as visual question answering and image captioning. Consequently, these models are now well-positioned to reason about the physical world, particularly within domains such as robotic manipulation. However, current VLMs are limited in their understanding of the physical concepts (e.g., material, fragility) of common objects, which restricts their usefulness for robotic manipulation tasks that involve interaction and physical reasoning about such objects. To address this limitation, we propose PhysObjects, an object-centric dataset of 39.6K crowd-sourced and 417K automated physical concept annotations of common household objects. We demonstrate that fine-tuning a VLM on PhysObjects improves its understanding of physical object concepts, including generalization to held-out concepts, by capturing human priors of these concepts from visual appearance. We incorporate this physically grounded VLM in an interactive framework with a large language model-based robotic planner, and show improved planning performance on tasks that require reasoning about physical object concepts, compared to baselines that do not leverage physically grounded VLMs. We additionally illustrate the benefits of our physically grounded VLM on a real robot, where it improves task success rates. We release our dataset and provide further details and visualizations of our results at https://iliad.stanford.edu/pg-vlm/.

Shromona Ghosh, Dorsa Sadigh, Pierluigi Nuzzo et al.

We address the problem of diagnosing and repairing specifications for hybrid systems formalized in signal temporal logic (STL). Our focus is on the setting of automatic synthesis of controllers in a model predictive control (MPC) framework. We build on recent approaches that reduce the controller synthesis problem to solving one or more mixed integer linear programs (MILPs), where infeasibility of a MILP usually indicates unrealizability of the controller synthesis problem. Given an infeasible STL synthesis problem, we present algorithms that provide feedback on the reasons for unrealizability, and suggestions for making it realizable. Our algorithms are sound and complete, i.e., they provide a correct diagnosis, and always terminate with a non-trivial specification that is feasible using the chosen synthesis method, when such a solution exists. We demonstrate the effectiveness of our approach on the synthesis of controllers for various cyber-physical systems, including an autonomous driving application and an aircraft electric power system.

Joey Hejna, Dorsa Sadigh

While reinforcement learning (RL) has become a more popular approach for robotics, designing sufficiently informative reward functions for complex tasks has proven to be extremely difficult due their inability to capture human intent and policy exploitation. Preference based RL algorithms seek to overcome these challenges by directly learning reward functions from human feedback. Unfortunately, prior work either requires an unreasonable number of queries implausible for any human to answer or overly restricts the class of reward functions to guarantee the elicitation of the most informative queries, resulting in models that are insufficiently expressive for realistic robotics tasks. Contrary to most works that focus on query selection to \emph{minimize} the amount of data required for learning reward functions, we take an opposite approach: \emph{expanding} the pool of available data by viewing human-in-the-loop RL through the more flexible lens of multi-task learning. Motivated by the success of meta-learning, we pre-train preference models on prior task data and quickly adapt them for new tasks using only a handful of queries. Empirically, we reduce the amount of online feedback needed to train manipulation policies in Meta-World by 20$\times$, and demonstrate the effectiveness of our method on a real Franka Panda Robot. Moreover, this reduction in query-complexity allows us to train robot policies from actual human users. Videos of our results and code can be found at https://sites.google.com/view/few-shot-preference-rl/home.

Sumedh A Sontakke, Jesse Zhang, Sébastien M. R. Arnold et al.

Reward specification is a notoriously difficult problem in reinforcement learning, requiring extensive expert supervision to design robust reward functions. Imitation learning (IL) methods attempt to circumvent these problems by utilizing expert demonstrations but typically require a large number of in-domain expert demonstrations. Inspired by advances in the field of Video-and-Language Models (VLMs), we present RoboCLIP, an online imitation learning method that uses a single demonstration (overcoming the large data requirement) in the form of a video demonstration or a textual description of the task to generate rewards without manual reward function design. Additionally, RoboCLIP can also utilize out-of-domain demonstrations, like videos of humans solving the task for reward generation, circumventing the need to have the same demonstration and deployment domains. RoboCLIP utilizes pretrained VLMs without any finetuning for reward generation. Reinforcement learning agents trained with RoboCLIP rewards demonstrate 2-3 times higher zero-shot performance than competing imitation learning methods on downstream robot manipulation tasks, doing so using only one video/text demonstration.

Andy Shih, Dorsa Sadigh, Stefano Ermon

Conditional inference on arbitrary subsets of variables is a core problem in probabilistic inference with important applications such as masked language modeling and image inpainting. In recent years, the family of Any-Order Autoregressive Models (AO-ARMs) -- closely related to popular models such as BERT and XLNet -- has shown breakthrough performance in arbitrary conditional tasks across a sweeping range of domains. But, in spite of their success, in this paper we identify significant improvements to be made to previous formulations of AO-ARMs. First, we show that AO-ARMs suffer from redundancy in their probabilistic model, i.e., they define the same distribution in multiple different ways. We alleviate this redundancy by training on a smaller set of univariate conditionals that still maintains support for efficient arbitrary conditional inference. Second, we upweight the training loss for univariate conditionals that are evaluated more frequently during inference. Our method leads to improved performance with no compromises on tractability, giving state-of-the-art likelihoods in arbitrary conditional modeling on text (Text8), image (CIFAR10, ImageNet32), and continuous tabular data domains.

Jonathan Yang, Dorsa Sadigh, Chelsea Finn

Reusing large datasets is crucial to scale vision-based robotic manipulators to everyday scenarios due to the high cost of collecting robotic datasets. However, robotic platforms possess varying control schemes, camera viewpoints, kinematic configurations, and end-effector morphologies, posing significant challenges when transferring manipulation skills from one platform to another. To tackle this problem, we propose a set of key design decisions to train a single policy for deployment on multiple robotic platforms. Our framework first aligns the observation and action spaces of our policy across embodiments via utilizing wrist cameras and a unified, but modular codebase. To bridge the remaining domain shift, we align our policy's internal representations across embodiments through contrastive learning. We evaluate our method on a dataset collected over 60 hours spanning 6 tasks and 3 robots with varying joint configurations and sizes: the WidowX 250S, the Franka Emika Panda, and the Sawyer. Our results demonstrate significant improvements in success rate and sample efficiency for our policy when using new task data collected on a different robot, validating our proposed design decisions. More details and videos can be found on our anonymized project website: https://sites.google.com/view/polybot-multirobot

Homanga Bharadhwaj, Debidatta Dwibedi, Abhinav Gupta et al.

How can robot manipulation policies generalize to novel tasks involving unseen object types and new motions? In this paper, we provide a solution in terms of predicting motion information from web data through human video generation and conditioning a robot policy on the generated video. Instead of attempting to scale robot data collection which is expensive, we show how we can leverage video generation models trained on easily available web data, for enabling generalization. Our approach Gen2Act casts language-conditioned manipulation as zero-shot human video generation followed by execution with a single policy conditioned on the generated video. To train the policy, we use an order of magnitude less robot interaction data compared to what the video prediction model was trained on. Gen2Act doesn't require fine-tuning the video model at all and we directly use a pre-trained model for generating human videos. Our results on diverse real-world scenarios show how Gen2Act enables manipulating unseen object types and performing novel motions for tasks not present in the robot data. Videos are at https://homangab.github.io/gen2act/

Hengyuan Hu, Dorsa Sadigh

One of the fundamental quests of AI is to produce agents that coordinate well with humans. This problem is challenging, especially in domains that lack high quality human behavioral data, because multi-agent reinforcement learning (RL) often converges to different equilibria from the ones that humans prefer. We propose a novel framework, instructRL, that enables humans to specify what kind of strategies they expect from their AI partners through natural language instructions. We use pretrained large language models to generate a prior policy conditioned on the human instruction and use the prior to regularize the RL objective. This leads to the RL agent converging to equilibria that are aligned with human preferences. We show that instructRL converges to human-like policies that satisfy the given instructions in a proof-of-concept environment as well as the challenging Hanabi benchmark. Finally, we show that knowing the language instruction significantly boosts human-AI coordination performance in human evaluations in Hanabi.

Suneel Belkhale, Yuchen Cui, Dorsa Sadigh

In supervised learning, the question of data quality and curation has been over-shadowed in recent years by increasingly more powerful and expressive models that can ingest internet-scale data. However, in offline learning for robotics, we simply lack internet scale data, and so high quality datasets are a necessity. This is especially true in imitation learning (IL), a sample efficient paradigm for robot learning using expert demonstrations. Policies learned through IL suffer from state distribution shift at test time due to compounding errors in action prediction, which leads to unseen states that the policy cannot recover from. Instead of designing new algorithms to address distribution shift, an alternative perspective is to develop new ways of assessing and curating datasets. There is growing evidence that the same IL algorithms can have substantially different performance across different datasets. This calls for a formalism for defining metrics of "data quality" that can further be leveraged for data curation. In this work, we take the first step toward formalizing data quality for imitation learning through the lens of distribution shift: a high quality dataset encourages the policy to stay in distribution at test time. We propose two fundamental properties that shape the quality of a dataset: i) action divergence: the mismatch between the expert and learned policy at certain states; and ii) transition diversity: the noise present in the system for a given state and action. We investigate the combined effect of these two key properties in imitation learning theoretically, and we empirically analyze models trained on a variety of different data sources. We show that state diversity is not always beneficial, and we demonstrate how action divergence and transition diversity interact in practice.

Joey Hejna, Rafael Rafailov, Harshit Sikchi et al.

Reinforcement Learning from Human Feedback (RLHF) has emerged as a popular paradigm for aligning models with human intent. Typically RLHF algorithms operate in two phases: first, use human preferences to learn a reward function and second, align the model by optimizing the learned reward via reinforcement learning (RL). This paradigm assumes that human preferences are distributed according to reward, but recent work suggests that they instead follow the regret under the user's optimal policy. Thus, learning a reward function from feedback is not only based on a flawed assumption of human preference, but also leads to unwieldy optimization challenges that stem from policy gradients or bootstrapping in the RL phase. Because of these optimization challenges, contemporary RLHF methods restrict themselves to contextual bandit settings (e.g., as in large language models) or limit observation dimensionality (e.g., state-based robotics). We overcome these limitations by introducing a new family of algorithms for optimizing behavior from human feedback using the regret-based model of human preferences. Using the principle of maximum entropy, we derive Contrastive Preference Learning (CPL), an algorithm for learning optimal policies from preferences without learning reward functions, circumventing the need for RL. CPL is fully off-policy, uses only a simple contrastive objective, and can be applied to arbitrary MDPs. This enables CPL to elegantly scale to high-dimensional and sequential RLHF problems while being simpler than prior methods.

Lihan Zha, Yuchen Cui, Li-Heng Lin et al.

Today's robot policies exhibit subpar performance when faced with the challenge of generalizing to novel environments. Human corrective feedback is a crucial form of guidance to enable such generalization. However, adapting to and learning from online human corrections is a non-trivial endeavor: not only do robots need to remember human feedback over time to retrieve the right information in new settings and reduce the intervention rate, but also they would need to be able to respond to feedback that can be arbitrary corrections about high-level human preferences to low-level adjustments to skill parameters. In this work, we present Distillation and Retrieval of Online Corrections (DROC), a large language model (LLM)-based system that can respond to arbitrary forms of language feedback, distill generalizable knowledge from corrections, and retrieve relevant past experiences based on textual and visual similarity for improving performance in novel settings. DROC is able to respond to a sequence of online language corrections that address failures in both high-level task plans and low-level skill primitives. We demonstrate that DROC effectively distills the relevant information from the sequence of online corrections in a knowledge base and retrieves that knowledge in settings with new task or object instances. DROC outperforms other techniques that directly generate robot code via LLMs by using only half of the total number of corrections needed in the first round and requires little to no corrections after two iterations. We show further results, videos, prompts and code on https://sites.google.com/stanford.edu/droc .

Andy Shih, Dorsa Sadigh, Stefano Ermon

Temperature scaling is a popular technique for tuning the sharpness of a model distribution. It is used extensively for sampling likely generations and calibrating model uncertainty, and even features as a controllable parameter to many large language models in deployment. However, autoregressive models rely on myopic temperature scaling that greedily optimizes the next token. To address this, we propose Long Horizon Temperature Scaling (LHTS), a novel approach for sampling from temperature-scaled joint distributions. LHTS is compatible with all likelihood-based models, and optimizes for the long horizon likelihood of samples. We derive a temperature-dependent LHTS objective, and show that finetuning a model on a range of temperatures produces a single model capable of generation with a controllable long horizon temperature parameter. We experiment with LHTS on image diffusion models and character/language autoregressive models, demonstrating advantages over myopic temperature scaling in likelihood and sample quality, and showing improvements in accuracy on a multiple choice analogy task by $10\%$.

Jennifer Grannen, Yilin Wu, Suneel Belkhale et al.

A robotic feeding system must be able to acquire a variety of foods. Prior bite acquisition works consider single-arm spoon scooping or fork skewering, which do not generalize to foods with complex geometries and deformabilities. For example, when acquiring a group of peas, skewering could smoosh the peas while scooping without a barrier could result in chasing the peas on the plate. In order to acquire foods with such diverse properties, we propose stabilizing food items during scooping using a second arm, for example, by pushing peas against the spoon with a flat surface to prevent dispersion. The added stabilizing arm can lead to new challenges. Critically, this arm should stabilize the food scene without interfering with the acquisition motion, which is especially difficult for easily breakable high-risk food items like tofu. These high-risk foods can break between the pusher and spoon during scooping, which can lead to food waste falling out of the spoon. We propose a general bimanual scooping primitive and an adaptive stabilization strategy that enables successful acquisition of a diverse set of food geometries and physical properties. Our approach, CARBS: Coordinated Acquisition with Reactive Bimanual Scooping, learns to stabilize without impeding task progress by identifying high-risk foods and robustly scooping them using closed-loop visual feedback. We find that CARBS is able to generalize across food shape, size, and deformability and is additionally able to manipulate multiple food items simultaneously. CARBS achieves 87.0% success on scooping rigid foods, which is 25.8% more successful than a single-arm baseline, and reduces food breakage by 16.2% compared to an analytical baseline. Videos can be found at https://sites.google.com/view/bimanualscoop-corl22/home .

Priya Sundaresan, Suneel Belkhale, Dorsa Sadigh

Acquiring food items with a fork poses an immense challenge to a robot-assisted feeding system, due to the wide range of material properties and visual appearances present across food groups. Deformable foods necessitate different skewering strategies than firm ones, but inferring such characteristics for several previously unseen items on a plate remains nontrivial. Our key insight is to leverage visual and haptic observations during interaction with an item to rapidly and reactively plan skewering motions. We learn a generalizable, multimodal representation for a food item from raw sensory inputs which informs the optimal skewering strategy. Given this representation, we propose a zero-shot framework to sense visuo-haptic properties of a previously unseen item and reactively skewer it, all within a single interaction. Real-robot experiments with foods of varying levels of visual and textural diversity demonstrate that our multimodal policy outperforms baselines which do not exploit both visual and haptic cues or do not reactively plan. Across 6 plates of different food items, our proposed framework achieves 71% success over 69 skewering attempts total. Supplementary material, datasets, code, and videos are available on our website: https://sites.google.com/view/hapticvisualnet-corl22/home

Tim Salzmann, Lewis Chiang, Markus Ryll et al.

Anticipating the motion of all humans in dynamic environments such as homes and offices is critical to enable safe and effective robot navigation. Such spaces remain challenging as humans do not follow strict rules of motion and there are often multiple occluded entry points such as corners and doors that create opportunities for sudden encounters. In this work, we present a Transformer based architecture to predict human future trajectories in human-centric environments from input features including human positions, head orientations, and 3D skeletal keypoints from onboard in-the-wild sensory information. The resulting model captures the inherent uncertainty for future human trajectory prediction and achieves state-of-the-art performance on common prediction benchmarks and a human tracking dataset captured from a mobile robot adapted for the prediction task. Furthermore, we identify new agents with limited historical data as a major contributor to error and demonstrate the complementary nature of 3D skeletal poses in reducing prediction error in such challenging scenarios.

Suvir Mirchandani, Mia Tang, Jiafei Duan et al.

Imitation learning from human demonstrations has become a dominant approach for training autonomous robot policies. However, collecting demonstration datasets is costly: it often requires access to robots and needs sustained effort in a tedious, long process. These factors limit the scale of data available for training policies. We aim to address this scalability challenge by involving a broader audience in a gamified data collection experience that is both accessible and motivating. Specifically, we develop a gamified remote teleoperation platform, RoboCade, to engage general users in collecting data that is beneficial for downstream policy training. To do this, we embed gamification strategies into the design of the system interface and data collection tasks. In the system interface, we include components such as visual feedback, sound effects, goal visualizations, progress bars, leaderboards, and badges. We additionally propose principles for constructing gamified tasks that have overlapping structure with useful downstream target tasks. We instantiate RoboCade on three manipulation tasks -- including spatial arrangement, scanning, and insertion. To illustrate the viability of gamified robot data collection, we collect a demonstration dataset through our platform, and show that co-training robot policies with this data can improve success rate on non-gamified target tasks (+16-56%). Further, we conduct a user study to validate that novice users find the gamified platform significantly more enjoyable than a standard non-gamified platform (+24%). These results highlight the promise of gamified data collection as a scalable, accessible, and engaging method for collecting demonstration data.

Perry Dong, Kuo-Han Hung, Tian Gao et al.

The ability to efficiently and reliably learn new tasks has been a foundational challenge in robotics. Vision-Language-Action (VLA) models have demonstrated strong generalization across diverse manipulation tasks, yet pretrained policies consistently fall short of the reliability required for real-world deployment. Reinforcement learning (RL) fine-tuning offers a promising path to bridge this gap, but existing approaches either train from scratch without fully leveraging pretrained priors, or fine-tune VLAs without achieving the sample efficiency and success rates that practical deployment demands. We present EXPO-FT, a system for stable, sample-efficient RL finetuning of pretrained VLA policies that closes this gap. Our system solves a suite of challenging manipulation tasks, including routing string lights and inserting the plug to light it up, striking a pool ball into a pocket, and inserting a flower into a wine bottle, each requiring combinations of high precision, dynamic actions, and robustness to varied initial states. Our system achieves perfect task performance (30/30 successes) across all evaluated tasks within an average of 19.1 minutes of online robot data, outperforming both prior RL-from-scratch and VLA finetuning approaches. We release an open-source codebase with the aim of facilitating broader adoption of RL finetuning of VLA models in robotics.

Priya Sundaresan, Suneel Belkhale, Dorsa Sadigh et al.

While natural language offers a convenient shared interface for humans and robots, enabling robots to interpret and follow language commands remains a longstanding challenge in manipulation. A crucial step to realizing a performant instruction-following robot is achieving semantic manipulation, where a robot interprets language at different specificities, from high-level instructions like "Pick up the stuffed animal" to more detailed inputs like "Grab the left ear of the elephant." To tackle this, we propose Keypoints + Instructions to Execution (KITE), a two-step framework for semantic manipulation which attends to both scene semantics (distinguishing between different objects in a visual scene) and object semantics (precisely localizing different parts within an object instance). KITE first grounds an input instruction in a visual scene through 2D image keypoints, providing a highly accurate object-centric bias for downstream action inference. Provided an RGB-D scene observation, KITE then executes a learned keypoint-conditioned skill to carry out the instruction. The combined precision of keypoints and parameterized skills enables fine-grained manipulation with generalization to scene and object variations. Empirically, we demonstrate KITE in 3 real-world environments: long-horizon 6-DoF tabletop manipulation, semantic grasping, and a high-precision coffee-making task. In these settings, KITE achieves a 75%, 70%, and 71% overall success rate for instruction-following, respectively. KITE outperforms frameworks that opt for pre-trained visual language models over keypoint-based grounding, or omit skills in favor of end-to-end visuomotor control, all while being trained from fewer or comparable amounts of demonstrations. Supplementary material, datasets, code, and videos can be found on our website: http://tinyurl.com/kite-site.

Li-Heng Lin, Yuchen Cui, Amber Xie et al.

Few-shot imitation learning relies on only a small amount of task-specific demonstrations to efficiently adapt a policy for a given downstream tasks. Retrieval-based methods come with a promise of retrieving relevant past experiences to augment this target data when learning policies. However, existing data retrieval methods fall under two extremes: they either rely on the existence of exact behaviors with visually similar scenes in the prior data, which is impractical to assume; or they retrieve based on semantic similarity of high-level language descriptions of the task, which might not be that informative about the shared low-level behaviors or motions across tasks that is often a more important factor for retrieving relevant data for policy learning. In this work, we investigate how we can leverage motion similarity in the vast amount of cross-task data to improve few-shot imitation learning of the target task. Our key insight is that motion-similar data carries rich information about the effects of actions and object interactions that can be leveraged during few-shot adaptation. We propose FlowRetrieval, an approach that leverages optical flow representations for both extracting similar motions to target tasks from prior data, and for guiding learning of a policy that can maximally benefit from such data. Our results show FlowRetrieval significantly outperforms prior methods across simulated and real-world domains, achieving on average 27% higher success rate than the best retrieval-based prior method. In the Pen-in-Cup task with a real Franka Emika robot, FlowRetrieval achieves 3.7x the performance of the baseline imitation learning technique that learns from all prior and target data. Website: https://flow-retrieval.github.io

Rajat Kumar Jenamani, Priya Sundaresan, Maram Sakr et al.

Robot-assisted feeding has the potential to improve the quality of life for individuals with mobility limitations who are unable to feed themselves independently. However, there exists a large gap between the homogeneous, curated plates existing feeding systems can handle, and truly in-the-wild meals. Feeding realistic plates is immensely challenging due to the sheer range of food items that a robot may encounter, each requiring specialized manipulation strategies which must be sequenced over a long horizon to feed an entire meal. An assistive feeding system should not only be able to sequence different strategies efficiently in order to feed an entire meal, but also be mindful of user preferences given the personalized nature of the task. We address this with FLAIR, a system for long-horizon feeding which leverages the commonsense and few-shot reasoning capabilities of foundation models, along with a library of parameterized skills, to plan and execute user-preferred and efficient bite sequences. In real-world evaluations across 6 realistic plates, we find that FLAIR can effectively tap into a varied library of skills for efficient food pickup, while adhering to the diverse preferences of 42 participants without mobility limitations as evaluated in a user study. We demonstrate the seamless integration of FLAIR with existing bite transfer methods [19, 28], and deploy it across 2 institutions and 3 robots, illustrating its adaptability. Finally, we illustrate the real-world efficacy of our system by successfully feeding a care recipient with severe mobility limitations. Supplementary materials and videos can be found at: https://emprise.cs.cornell.edu/flair .

Priya Sundaresan, Jiajun Wu, Dorsa Sadigh

A robot providing mealtime assistance must perform specialized maneuvers with various utensils in order to pick up and feed a range of food items. Beyond these dexterous low-level skills, an assistive robot must also plan these strategies in sequence over a long horizon to clear a plate and complete a meal. Previous methods in robot-assisted feeding introduce highly specialized primitives for food handling without a means to compose them together. Meanwhile, existing approaches to long-horizon manipulation lack the flexibility to embed highly specialized primitives into their frameworks. We propose Visual Action Planning OveR Sequences (VAPORS), a framework for long-horizon food acquisition. VAPORS learns a policy for high-level action selection by leveraging learned latent plate dynamics in simulation. To carry out sequential plans in the real world, VAPORS delegates action execution to visually parameterized primitives. We validate our approach on complex real-world acquisition trials involving noodle acquisition and bimanual scooping of jelly beans. Across 38 plates, VAPORS acquires much more efficiently than baselines, generalizes across realistic plate variations such as toppings and sauces, and qualitatively appeals to user feeding preferences in a survey conducted across 49 individuals. Code, datasets, videos, and supplementary materials can be found on our website: https://sites.google.com/view/vaporsbot.

Joey Hejna, Jensen Gao, Dorsa Sadigh

Sequential decision making algorithms often struggle to leverage different sources of unstructured offline interaction data. Imitation learning (IL) methods based on supervised learning are robust, but require optimal demonstrations, which are hard to collect. Offline goal-conditioned reinforcement learning (RL) algorithms promise to learn from sub-optimal data, but face optimization challenges especially with high-dimensional data. To bridge the gap between IL and RL, we introduce Distance Weighted Supervised Learning or DWSL, a supervised method for learning goal-conditioned policies from offline data. DWSL models the entire distribution of time-steps between states in offline data with only supervised learning, and uses this distribution to approximate shortest path distances. To extract a policy, we weight actions by their reduction in distance estimates. Theoretically, DWSL converges to an optimal policy constrained to the data distribution, an attractive property for offline learning, without any bootstrapping. Across all datasets we test, DWSL empirically maintains behavior cloning as a lower bound while still exhibiting policy improvement. In high-dimensional image domains, DWSL surpasses the performance of both prior goal-conditioned IL and RL algorithms. Visualizations and code can be found at https://sites.google.com/view/dwsl/home .

Bidipta Sarkar, Andy Shih, Dorsa Sadigh

Conventions are crucial for strong performance in cooperative multi-agent games, because they allow players to coordinate on a shared strategy without explicit communication. Unfortunately, standard multi-agent reinforcement learning techniques, such as self-play, converge to conventions that are arbitrary and non-diverse, leading to poor generalization when interacting with new partners. In this work, we present a technique for generating diverse conventions by (1) maximizing their rewards during self-play, while (2) minimizing their rewards when playing with previously discovered conventions (cross-play), stimulating conventions to be semantically different. To ensure that learned policies act in good faith despite the adversarial optimization of cross-play, we introduce \emph{mixed-play}, where an initial state is randomly generated by sampling self-play and cross-play transitions and the player learns to maximize the self-play reward from this initial state. We analyze the benefits of our technique on various multi-agent collaborative games, including Overcooked, and find that our technique can adapt to the conventions of humans, surpassing human-level performance when paired with real users.

Vivek Myers, Erdem Bıyık, Dorsa Sadigh

Robot policies need to adapt to human preferences and/or new environments. Human experts may have the domain knowledge required to help robots achieve this adaptation. However, existing works often require costly offline re-training on human feedback, and those feedback usually need to be frequent and too complex for the humans to reliably provide. To avoid placing undue burden on human experts and allow quick adaptation in critical real-world situations, we propose designing and sparingly presenting easy-to-answer pairwise action preference queries in an online fashion. Our approach designs queries and determines when to present them to maximize the expected value derived from the queries' information. We demonstrate our approach with experiments in simulation, human user studies, and real robot experiments. In these settings, our approach outperforms baseline techniques while presenting fewer queries to human experts. Experiment videos, code and appendices are found at https://sites.google.com/view/onlineactivepreferences.

Zhangjie Cao, Erdem Bıyık, Guy Rosman et al.

Multi-agent interactions are important to model for forecasting other agents' behaviors and trajectories. At a certain time, to forecast a reasonable future trajectory, each agent needs to pay attention to the interactions with only a small group of most relevant agents instead of unnecessarily paying attention to all the other agents. However, existing attention modeling works ignore that human attention in driving does not change rapidly, and may introduce fluctuating attention across time steps. In this paper, we formulate an attention model for multi-agent interactions based on a total variation temporal smoothness prior and propose a trajectory prediction architecture that leverages the knowledge of these attended interactions. We demonstrate how the total variation attention prior along with the new sequence prediction loss terms leads to smoother attention and more sample-efficient learning of multi-agent trajectory prediction, and show its advantages in terms of prediction accuracy by comparing it with the state-of-the-art approaches on both synthetic and naturalistic driving data. We demonstrate the performance of our algorithm for trajectory prediction on the INTERACTION dataset on our website.

Mengxi Li, Rika Antonova, Dorsa Sadigh et al.

When humans perform contact-rich manipulation tasks, customized tools are often necessary to simplify the task. For instance, we use various utensils for handling food, such as knives, forks and spoons. Similarly, robots may benefit from specialized tools that enable them to more easily complete a variety of tasks. We present an end-to-end framework to automatically learn tool morphology for contact-rich manipulation tasks by leveraging differentiable physics simulators. Previous work relied on manually constructed priors requiring detailed specification of a 3D object model, grasp pose and task description to facilitate the search or optimization process. Our approach only requires defining the objective with respect to task performance and enables learning a robust morphology through randomizing variations of the task. We make this optimization tractable by casting it as a continual learning problem. We demonstrate the effectiveness of our method for designing new tools in several scenarios, such as winding ropes, flipping a box and pushing peas onto a scoop in simulation. Additionally, experiments with real robots show that the tool shapes discovered by our method help them succeed in these scenarios.

Zihan Wang, Zhangjie Cao, Yilun Hao et al.

Correspondence learning is a fundamental problem in robotics, which aims to learn a mapping between state, action pairs of agents of different dynamics or embodiments. However, current correspondence learning methods either leverage strictly paired data -- which are often difficult to collect -- or learn in an unsupervised fashion from unpaired data using regularization techniques such as cycle-consistency -- which suffer from severe misalignment issues. We propose a weakly supervised correspondence learning approach that trades off between strong supervision over strictly paired data and unsupervised learning with a regularizer over unpaired data. Our idea is to leverage two types of weak supervision: i) temporal ordering of states and actions to reduce the compounding error, and ii) paired abstractions, instead of paired data, to alleviate the misalignment problem and learn a more accurate correspondence. The two types of weak supervision are easy to access in real-world applications, which simultaneously reduces the high cost of annotating strictly paired data and improves the quality of the learned correspondence.

Maximilian Du, Jennifer Grannen, Shuran Song et al.

The physicality of exercise makes the role of athletic trainers unique. Their physical presence allows them to guide a student through a motion, demonstrate an exercise, and give intuitive feedback. Robot quadrupeds are also embodied agents with robust agility and athleticism. In our work, we investigate whether a robot quadruped can serve as an effective and enjoyable personal trainer device. We focus on a case study of interval training for runners: a repetitive, long-horizon task where precision and consistency are important. To meet this challenge, we propose SNOOPIE, an autonomous robot quadruped pacer capable of running interval training exercises tailored to challenge a user's personal abilities. We conduct a set of user experiments that compare the robot trainer to a wearable trainer device--the Apple Watch--to investigate the benefits of a physical embodiment in exercise-based interactions. We demonstrate 60.6% better adherence to a pace schedule and were 45.9% more consistent across their running speeds with the quadruped trainer. Subjective results also showed that participants strongly preferred training with the robot over wearable devices across many qualitative axes, including its ease of use (+56.7%), enjoyability of the interaction (+60.6%), and helpfulness (+39.1%). Additional videos and visualizations can be found on our website: https://sites.google.com/view/snoopie

Megha Srivastava, Noah Goodman, Dorsa Sadigh

AI assistance continues to help advance applications in education, from language learning to intelligent tutoring systems, yet current methods for providing students feedback are still quite limited. Most automatic feedback systems either provide binary correctness feedback, which may not help a student understand how to improve, or require hand-coding feedback templates, which may not generalize to new domains. This can be particularly challenging for physical control tasks, where the rich diversity in student behavior and specialized domains make it challenging to leverage general-purpose assistive tools for providing feedback. We design and build CORGI, a model trained to generate language corrections for physical control tasks, such as learning to ride a bike. CORGI takes in as input a pair of student and expert trajectories, and then generates natural language corrections to help the student improve. We collect and train CORGI over data from three diverse physical control tasks (drawing, steering, and joint movement). Through both automatic and human evaluations, we show that CORGI can (i) generate valid feedback for novel student trajectories, (ii) outperform baselines on domains with novel control dynamics, and (iii) improve student learning in an interactive drawing task.

Ifdita Hasan Orney, Jubayer Ibn Hamid, Shreya S Ramanujam et al.

Exploration is a cornerstone of learning from experience: it enables agents to find solutions to complex problems, generalize to novel ones, and scale performance with test-time compute. In this paper, we present a framework for post-training language models (LMs) that explicitly encourages optimistic exploration and promotes a synergy between exploration and exploitation. The central idea is to train the LM to generate sets of responses that are collectively accurate under the reward function and exploratory in their reasoning strategies. We first develop a general recipe for optimizing LMs with set reinforcement learning (set RL) under arbitrary objective functions, showing how standard RL algorithms can be adapted to this setting through a modification to the advantage computation. We then propose Polychromic Exploratory Policy Optimization (Poly-EPO), which instantiates this framework with an objective that explicitly synergizes exploration and exploitation. Across a range of reasoning benchmarks, we show that Poly-EPO improves generalization, as evidenced by higher pass@$k$ coverage, preserves greater diversity in model generations, and effectively scales with test-time compute.

Yilun Hao, Ruinan Wang, Zhangjie Cao et al.

Multimodal demonstrations provide robots with an abundance of information to make sense of the world. However, such abundance may not always lead to good performance when it comes to learning sensorimotor control policies from human demonstrations. Extraneous data modalities can lead to state over-specification, where the state contains modalities that are not only useless for decision-making but also can change data distribution across environments. State over-specification leads to issues such as the learned policy not generalizing outside of the training data distribution. In this work, we propose Masked Imitation Learning (MIL) to address state over-specification by selectively using informative modalities. Specifically, we design a masked policy network with a binary mask to block certain modalities. We develop a bi-level optimization algorithm that learns this mask to accurately filter over-specified modalities. We demonstrate empirically that MIL outperforms baseline algorithms in simulated domains including MuJoCo and a robot arm environment using the Robomimic dataset, and effectively recovers the environment-invariant modalities on a multimodal dataset collected on a real robot. Our project website presents supplemental details and videos of our results at: https://tinyurl.com/masked-il

Octo Model Team, Dibya Ghosh, Homer Walke et al. · berkeley

Large policies pretrained on diverse robot datasets have the potential to transform robotic learning: instead of training new policies from scratch, such generalist robot policies may be finetuned with only a little in-domain data, yet generalize broadly. However, to be widely applicable across a range of robotic learning scenarios, environments, and tasks, such policies need to handle diverse sensors and action spaces, accommodate a variety of commonly used robotic platforms, and finetune readily and efficiently to new domains. In this work, we aim to lay the groundwork for developing open-source, widely applicable, generalist policies for robotic manipulation. As a first step, we introduce Octo, a large transformer-based policy trained on 800k trajectories from the Open X-Embodiment dataset, the largest robot manipulation dataset to date. It can be instructed via language commands or goal images and can be effectively finetuned to robot setups with new sensory inputs and action spaces within a few hours on standard consumer GPUs. In experiments across 9 robotic platforms, we demonstrate that Octo serves as a versatile policy initialization that can be effectively finetuned to new observation and action spaces. We also perform detailed ablations of design decisions for the Octo model, from architecture to training data, to guide future research on building generalist robot models.

Amber Xie, Haozhi Qi, Dorsa Sadigh

Mastering dexterous manipulation with multi-fingered hands has been a grand challenge in robotics for decades. Despite its potential, the difficulty of collecting high-quality data remains a primary bottleneck for high-precision tasks. While reinforcement learning and simulation-to-real-world transfer offer a promising alternative, the transferred policies often fail for tasks demanding millimeter-scale precision, such as bimanual piano playing. In this work, we introduce HandelBot, a framework that combines a simulation policy and rapid adaptation through a two-stage pipeline. Starting from a simulation-trained policy, we first apply a structured refinement stage to correct spatial alignments by adjusting lateral finger joints based on physical rollouts. Next, we use residual reinforcement learning to autonomously learn fine-grained corrective actions. Through extensive hardware experiments across five recognized songs, we demonstrate that HandelBot can successfully perform precise bimanual piano playing. Our system outperforms direct simulation deployment by a factor of 1.8x and requires only 30 minutes of physical interaction data.

Hung-Chieh Fang, Amber Xie, Jennifer Grannen et al.

Performing in-hand, contact-rich, and long-horizon dexterous manipulation remains an unsolved challenge in robotics. Prior hand dexterity works have considered each of these three challenges in isolation, yet do not combine these skills into a single, complex task. To further test the capabilities of dexterity, we propose drumming as a testbed for dexterous manipulation. Drumming naturally integrates all three challenges: it involves in-hand control for stabilizing and adjusting the drumstick with the fingers, contact-rich interaction through repeated striking of the drum surface, and long-horizon coordination when switching between drums and sustaining rhythmic play. We present DexDrummer, a hierarchical object-centric bimanual drumming policy trained in simulation with sim-to-real transfer. The framework reduces the exploration difficulty of pure reinforcement learning by combining trajectory planning with residual RL corrections for fast transitions between drums. A dexterous manipulation policy handles contact-rich dynamics, guided by rewards that explicitly model both finger-stick and stick-drum interactions. In simulation, we show our policy can play two styles of music: multi-drum, bimanual songs and challenging, technical exercises that require increased dexterity. Across simulated bimanual tasks, our dexterous, reactive policy outperforms a fixed grasp policy by 1.87x across easy songs and 1.22x across hard songs F1 scores. In real-world tasks, we show song performance across a multi-drum setup. DexDrummer is able to play our training song and its extended version with an F1 score of 1.0.

Jensen Gao, Dorsa Sadigh, Sandy Huang et al.

Recent work on robot manipulation has advanced policy generalization to novel scenarios. However, it is often difficult to characterize how different evaluation settings actually represent generalization from the training distribution of a given policy. To work towards more precise evaluation of generalization in robotics, we propose RADAR, a scalable framework for directly comparing test-time evaluation tasks to policy training data, to determine what form of policy generalization is required. RADAR consists of a two-stage pipeline: first, retrieval using generalist policy embeddings identifies which training examples are relevant for a given evaluation task. Next, vision-language models (VLMs) analyze the evaluation task against the retrieved data, outputting interpretable analysis on how they compare along a variety of axes, and an overall classification of what type of policy generalization is required. Through controlled experiments, we demonstrate that VLMs are effective at analyzing data for generalization, and that our retrieval step effectively identifies examples needed to make accurate classifications with respect to the training data. Furthermore, we scale RADAR to large-scale datasets, where we observe agreement with human-defined benchmark conditions from prior work. We provide demonstrations at radar-analysis.github.io.