Joey Hejna

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.

Joey Hejna, Pieter Abbeel, Lerrel Pinto

Complex, long-horizon planning and its combinatorial nature pose steep challenges for learning-based agents. Difficulties in such settings are exacerbated in low data regimes where over-fitting stifles generalization and compounding errors hurt accuracy. In this work, we explore the use of an often unused source of auxiliary supervision: language. Inspired by recent advances in transformer-based models, we train agents with an instruction prediction loss that encourages learning temporally extended representations that operate at a high level of abstraction. Concretely, we demonstrate that instruction modeling significantly improves performance in planning environments when training with a limited number of demonstrations on the BabyAI and Crafter benchmarks. In further analysis we find that instruction modeling is most important for tasks that require complex reasoning, while understandably offering smaller gains in environments that require simple plans. More details and code can be found at https://github.com/jhejna/instruction-prediction.

Cursor Research, Aaron Chan, Ahmed Shalaby et al. · berkeley, microsoft-research

Composer 2 is a specialized model designed for agentic software engineering. The model demonstrates strong long-term planning and coding intelligence while maintaining the ability to efficiently solve problems for interactive use. The model is trained in two phases: first, continued pretraining to improve the model's knowledge and latent coding ability, followed by large-scale reinforcement learning to improve end-to-end coding performance through stronger reasoning, accurate multi-step execution, and coherence on long-horizon realistic coding problems. We develop infrastructure to support training in the same Cursor harness that is used by the deployed model, with equivalent tools and structure, and use environments that match real problems closely. To measure the ability of the model on increasingly difficult tasks, we introduce a benchmark derived from real software engineering problems in large codebases including our own. Composer 2 is a frontier-level coding model and demonstrates a process for training strong domain-specialized models. On our CursorBench evaluations the model achieves a major improvement in accuracy compared to previous Composer models (61.3). On public benchmarks the model scores 61.7 on Terminal-Bench and 73.7 on SWE-bench Multilingual in our harness, comparable to state-of-the-art systems.

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

Divyansh Garg, Joey Hejna, Matthieu Geist et al.

Modern Deep Reinforcement Learning (RL) algorithms require estimates of the maximal Q-value, which are difficult to compute in continuous domains with an infinite number of possible actions. In this work, we introduce a new update rule for online and offline RL which directly models the maximal value using Extreme Value Theory (EVT), drawing inspiration from economics. By doing so, we avoid computing Q-values using out-of-distribution actions which is often a substantial source of error. Our key insight is to introduce an objective that directly estimates the optimal soft-value functions (LogSumExp) in the maximum entropy RL setting without needing to sample from a policy. Using EVT, we derive our \emph{Extreme Q-Learning} framework and consequently online and, for the first time, offline MaxEnt Q-learning algorithms, that do not explicitly require access to a policy or its entropy. Our method obtains consistently strong performance in the D4RL benchmark, outperforming prior works by \emph{10+ points} on the challenging Franka Kitchen tasks while offering moderate improvements over SAC and TD3 on online DM Control tasks. Visualizations and code can be found on our website at https://div99.github.io/XQL/.

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.

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.

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 .

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.

Rafael Rafailov, Joey Hejna, Ryan Park et al.

Reinforcement Learning From Human Feedback (RLHF) has been critical to the success of the latest generation of generative AI models. In response to the complex nature of the classical RLHF pipeline, direct alignment algorithms such as Direct Preference Optimization (DPO) have emerged as an alternative approach. Although DPO solves the same objective as the standard RLHF setup, there is a mismatch between the two approaches. Standard RLHF deploys reinforcement learning in a specific token-level MDP, while DPO is derived as a bandit problem in which the whole response of the model is treated as a single arm. In this work we rectify this difference. We theoretically show that we can derive DPO in the token-level MDP as a general inverse Q-learning algorithm, which satisfies the Bellman equation. Using our theoretical results, we provide three concrete empirical insights. First, we show that because of its token level interpretation, DPO is able to perform some type of credit assignment. Next, we prove that under the token level formulation, classical search-based algorithms, such as MCTS, which have recently been applied to the language generation space, are equivalent to likelihood-based search on a DPO policy. Empirically we show that a simple beam search yields meaningful improvement over the base DPO policy. Finally, we show how the choice of reference policy causes implicit rewards to decline during training. We conclude by discussing applications of our work, including information elicitation in multi-turn dialogue, reasoning, agentic applications and end-to-end training of multi-model systems.

Yecheng Jason Ma, Joey Hejna, Ayzaan Wahid et al.

Predicting temporal progress from visual trajectories is important for intelligent robots that can learn, adapt, and improve. However, learning such progress estimator, or temporal value function, across different tasks and domains requires both a large amount of diverse data and methods which can scale and generalize. To address these challenges, we present Generative Value Learning (\GVL), a universal value function estimator that leverages the world knowledge embedded in vision-language models (VLMs) to predict task progress. Naively asking a VLM to predict values for a video sequence performs poorly due to the strong temporal correlation between successive frames. Instead, GVL poses value estimation as a temporal ordering problem over shuffled video frames; this seemingly more challenging task encourages VLMs to more fully exploit their underlying semantic and temporal grounding capabilities to differentiate frames based on their perceived task progress, consequently producing significantly better value predictions. Without any robot or task specific training, GVL can in-context zero-shot and few-shot predict effective values for more than 300 distinct real-world tasks across diverse robot platforms, including challenging bimanual manipulation tasks. Furthermore, we demonstrate that GVL permits flexible multi-modal in-context learning via examples from heterogeneous tasks and embodiments, such as human videos. The generality of GVL enables various downstream applications pertinent to visuomotor policy learning, including dataset filtering, success detection, and advantage-weighted regression -- all without any model training or finetuning.

Suvir Mirchandani, Suneel Belkhale, Joey Hejna et al. · stanford

A long-standing goal in robot learning is to develop methods for robots to acquire new skills autonomously. While reinforcement learning (RL) comes with the promise of enabling autonomous data collection, it remains challenging to scale in the real-world partly due to the significant effort required for environment design and instrumentation, including the need for designing reset functions or accurate success detectors. On the other hand, imitation learning (IL) methods require little to no environment design effort, but instead require significant human supervision in the form of collected demonstrations. To address these shortcomings, recent works in autonomous IL start with an initial seed dataset of human demonstrations that an autonomous policy can bootstrap from. While autonomous IL approaches come with the promise of addressing the challenges of autonomous RL as well as pure IL strategies, in this work, we posit that such techniques do not deliver on this promise and are still unable to scale up autonomous data collection in the real world. Through a series of real-world experiments, we demonstrate that these approaches, when scaled up to realistic settings, face much of the same scaling challenges as prior attempts in RL in terms of environment design. Further, we perform a rigorous study of autonomous IL methods across different data scales and 7 simulation and real-world tasks, and demonstrate that while autonomous data collection can modestly improve performance, simply collecting more human data often provides significantly more improvement. Our work suggests a negative result: that scaling up autonomous data collection for learning robot policies for real-world tasks is more challenging and impractical than what is suggested in prior work. We hope these insights about the core challenges of scaling up data collection help inform future efforts in autonomous learning.

Jaden Clark, Joey Hejna, Dorsa Sadigh

Expressive robotic behavior is essential for the widespread acceptance of robots in social environments. Recent advancements in learned legged locomotion controllers have enabled more dynamic and versatile robot behaviors. However, determining the optimal behavior for interactions with different users across varied scenarios remains a challenge. Current methods either rely on natural language input, which is efficient but low-resolution, or learn from human preferences, which, although high-resolution, is sample inefficient. This paper introduces a novel approach that leverages priors generated by pre-trained LLMs alongside the precision of preference learning. Our method, termed Language-Guided Preference Learning (LGPL), uses LLMs to generate initial behavior samples, which are then refined through preference-based feedback to learn behaviors that closely align with human expectations. Our core insight is that LLMs can guide the sampling process for preference learning, leading to a substantial improvement in sample efficiency. We demonstrate that LGPL can quickly learn accurate and expressive behaviors with as few as four queries, outperforming both purely language-parameterized models and traditional preference learning approaches. Website with videos: https://lgpl-gaits.github.io/

Amber Xie, Rahul Chand, Dorsa Sadigh et al.

While large-scale robot datasets have propelled recent progress in imitation learning, learning from smaller task specific datasets remains critical for deployment in new environments and unseen tasks. One such approach to few-shot imitation learning is retrieval-based imitation learning, which extracts relevant samples from large, widely available prior datasets to augment a limited demonstration dataset. To determine the relevant data from prior datasets, retrieval-based approaches most commonly calculate a prior data point's minimum distance to a point in the target dataset in latent space. While retrieval-based methods have shown success using this metric for data selection, we demonstrate its equivalence to the limit of a Gaussian kernel density (KDE) estimate of the target data distribution. This reveals two shortcomings of the retrieval rule used in prior work. First, it relies on high-variance nearest neighbor estimates that are susceptible to noise. Second, it does not account for the distribution of prior data when retrieving data. To address these issues, we introduce Importance Weighted Retrieval (IWR), which estimates importance weights, or the ratio between the target and prior data distributions for retrieval, using Gaussian KDEs. By considering the probability ratio, IWR seeks to mitigate the bias of previous selection rules, and by using reasonable modeling parameters, IWR effectively smooths estimates using all data points. Across both simulation environments and real-world evaluations on the Bridge dataset we find that our method, IWR, consistently improves performance of existing retrieval-based methods, despite only requiring minor modifications.

Rafael Rafailov, Yaswanth Chittepu, Ryan Park et al.

Reinforcement Learning from Human Feedback (RLHF) has been crucial to the recent success of Large Language Models (LLMs), however, it is often a complex and brittle process. In the classical RLHF framework, a reward model is first trained to represent human preferences, which is in turn used by an online reinforcement learning (RL) algorithm to optimize the LLM. A prominent issue with such methods is reward over-optimization or reward hacking, where performance as measured by the learned proxy reward model increases, but true quality plateaus or even deteriorates. Direct Alignment Algorithms (DDAs) like Direct Preference Optimization have emerged as alternatives to the classical RLHF pipeline by circumventing the reward modeling phase. However, although DAAs do not use a separate proxy reward model, they still commonly deteriorate from over-optimization. While the so-called reward hacking phenomenon is not well-defined for DAAs, we still uncover similar trends: at higher KL budgets, DAA algorithms exhibit similar degradation patterns to their classic RLHF counterparts. In particular, we find that DAA methods deteriorate not only across a wide range of KL budgets but also often before even a single epoch of the dataset is completed. Through extensive empirical experimentation, this work formulates and formalizes the reward over-optimization or hacking problem for DAAs and explores its consequences across objectives, training regimes, and model scales.

Omar Shaikh, Michelle S. Lam, Joey Hejna et al.

Language models are aligned to emulate the collective voice of many, resulting in outputs that align with no one in particular. Steering LLMs away from generic output is possible through supervised finetuning or RLHF, but requires prohibitively large datasets for new ad-hoc tasks. We argue that it is instead possible to align an LLM to a specific setting by leveraging a very small number (< 10) of demonstrations as feedback. Our method, Demonstration ITerated Task Optimization (DITTO), directly aligns language model outputs to a user's demonstrated behaviors. Derived using ideas from online imitation learning, DITTO cheaply generates online comparison data by treating users' demonstrations as preferred over output from the LLM and its intermediate checkpoints. Concretely, DITTO operates by having an LLM generate examples that are presumed to be inferior to expert demonstrations. The method iteratively constructs pairwise preference relationships between these LLM-generated samples and expert demonstrations, potentially including comparisons between different training checkpoints. These constructed preference pairs are then used to train the model using a preference optimization algorithm (e.g. DPO). We evaluate DITTO's ability to learn fine-grained style and task alignment across domains such as news articles, emails, and blog posts. Additionally, we conduct a user study soliciting a range of demonstrations from participants (N = 16). Across our benchmarks and user study, we find that win-rates for DITTO outperform few-shot prompting, supervised fine-tuning, and other self-play methods by an avg. of 19% points. By using demonstrations as feedback directly, DITTO offers a novel method for effective customization of LLMs.

Joey Hejna, Dorsa Sadigh

Reward functions are difficult to design and often hard to align with human intent. Preference-based Reinforcement Learning (RL) algorithms address these problems by learning reward functions from human feedback. However, the majority of preference-based RL methods naïvely combine supervised reward models with off-the-shelf RL algorithms. Contemporary approaches have sought to improve performance and query complexity by using larger and more complex reward architectures such as transformers. Instead of using highly complex architectures, we develop a new and parameter-efficient algorithm, Inverse Preference Learning (IPL), specifically designed for learning from offline preference data. Our key insight is that for a fixed policy, the $Q$-function encodes all information about the reward function, effectively making them interchangeable. Using this insight, we completely eliminate the need for a learned reward function. Our resulting algorithm is simpler and more parameter-efficient. Across a suite of continuous control and robotics benchmarks, IPL attains competitive performance compared to more complex approaches that leverage transformer-based and non-Markovian reward functions while having fewer algorithmic hyperparameters and learned network parameters. Our code is publicly released.