Automated poultry processing lines still rely on humans to lift slippery, easily bruised carcasses onto a shackle conveyor. Deformability, anatomical variance, and strict hygiene rules make conventional suction and scripted motions unreliable. We present ChicGrasp, an end--to--end hardware--software co-design for this task. An independently actuated dual-jaw pneumatic gripper clamps both chicken legs, while a conditional diffusion-policy controller, trained from only 50 multi--view teleoperation demonstrations (RGB + proprioception), plans 5 DoF end--effector motion, which includes jaw commands in one shot. On individually presented raw broiler carcasses, our system achieves a 40.6\% grasp--and--lift success rate and completes the pick to shackle cycle in 38 s, whereas state--of--the--art implicit behaviour cloning (IBC) and LSTM-GMM baselines fail entirely. All CAD, code, and datasets will be open-source. ChicGrasp shows that imitation learning can bridge the gap between rigid hardware and variable bio--products, offering a reproducible benchmark and a public dataset for researchers in agricultural engineering and robot learning.
Due to the deformability of garments, generating a large amount of high-quality data for robotic garment manipulation tasks is highly challenging. In this paper, we present a synthetic garment dataset that can be used for robotic garment folding. We begin by constructing geometric garment templates based on keypoints and applying generative models to generate realistic texture patterns. Leveraging these keypoint annotations, we generate folding demonstrations in simulation and train folding policies via closed-loop imitation learning. To improve robustness, we propose KG-DAgger, which uses a keypoint-based strategy to generate demonstration data for recovering from failures. KG-DAgger significantly improves the model performance, boosting the real-world success rate by 25\%. After training with 15K trajectories (about 2M image-action pairs), the model achieves a 75\% success rate in the real world. Experiments in both simulation and real-world settings validate the effectiveness of our proposed framework.
Reasoning over long sequences of observations and actions is essential for many robotic tasks. Yet, learning effective long-context policies from demonstrations remains challenging. As context length increases, training becomes increasingly expensive due to rising memory demands, and policy performance often degrades as a result of spurious correlations. Recent methods typically sidestep these issues by truncating context length, discarding historical information that may be critical for subsequent decisions. In this paper, we propose an alternative approach that explicitly regularizes the retention of past information. We first revisit the copycat problem in imitation learning and identify an opposite challenge in recent diffusion policies: rather than over-relying on prior actions, they often fail to capture essential dependencies between past and future actions. To address this, we introduce Past-Token Prediction (PTP), an auxiliary task in which the policy learns to predict past action tokens alongside future ones. This regularization significantly improves temporal modeling in the policy head, with minimal reliance on visual representations. Building on this observation, we further introduce a multistage training strategy: pre-train the visual encoder with short contexts, and fine-tune the policy head using cached long-context embeddings. This strategy preserves the benefits of PTP while greatly reducing memory and computational overhead. Finally, we extend PTP into a self-verification mechanism at test time, enabling the policy to score and select candidates consistent with past actions during inference. Experiments across four real-world and six simulated tasks demonstrate that our proposed method improves the performance of long-context diffusion policies by 3x and accelerates policy training by more than 10x.
The ability to learn from large batches of autonomously collected data for policy improvement -- a paradigm we refer to as batch online reinforcement learning -- holds the promise of enabling truly scalable robot learning by significantly reducing the need for human effort of data collection while getting benefits from self-improvement. Yet, despite the promise of this paradigm, it remains challenging to achieve due to algorithms not being able to learn effectively from the autonomous data. For example, prior works have applied imitation learning and filtered imitation learning methods to the batch online RL problem, but these algorithms often fail to efficiently improve from the autonomously collected data or converge quickly to a suboptimal point. This raises the question of what matters for effective batch online RL in robotics. Motivated by this question, we perform a systematic empirical study of three axes -- (i) algorithm class, (ii) policy extraction methods, and (iii) policy expressivity -- and analyze how these axes affect performance and scaling with the amount of autonomous data. Through our analysis, we make several observations. First, we observe that the use of Q-functions to guide batch online RL significantly improves performance over imitation-based methods. Building on this, we show that an implicit method of policy extraction -- via choosing the best action in the distribution of the policy -- is necessary over traditional policy extraction methods from offline RL. Next, we show that an expressive policy class is preferred over less expressive policy classes. Based on this analysis, we propose a general recipe for effective batch online RL. We then show a simple addition to the recipe of using temporally-correlated noise to obtain more diversity results in further performance gains. Our recipe obtains significantly better performance and scaling compared to prior methods.
Despite impressive progress in areas like mathematical reasoning, large language models still face significant challenges in consistently solving complex problems. Drawing inspiration from key human learning strategies, we propose two novel strategies to enhance the capability of large language models to solve these complex problems. First, Adaptive Difficulty Curriculum Learning (ADCL) is a novel curriculum learning strategy that tackles the Difficulty Shift phenomenon (i.e., a model's perception of problem difficulty dynamically changes during training) by periodically re-estimating difficulty within upcoming data batches to maintain alignment with the model's evolving capabilities. Second, Expert-Guided Self-Reformulation (EGSR) is a novel reinforcement learning strategy that bridges the gap between imitation learning and pure exploration by guiding models to reformulate expert solutions within their own conceptual framework, rather than relying on direct imitation, fostering deeper understanding and knowledge assimilation. Extensive experiments on challenging mathematical reasoning benchmarks, using Qwen2.5-7B as the base model, demonstrate that these human-inspired strategies synergistically and significantly enhance performance. Notably, their combined application improves performance over the standard Zero-RL baseline by 10% on the AIME24 benchmark and 16.6% on AIME25.
Human videos offer a scalable way to train robot manipulation policies, but lack the action labels needed by standard imitation learning algorithms. Existing cross-embodiment approaches try to map human motion to robot actions, but often fail when the embodiments differ significantly. We propose X-Sim, a real-to-sim-to-real framework that uses object motion as a dense and transferable signal for learning robot policies. X-Sim starts by reconstructing a photorealistic simulation from an RGBD human video and tracking object trajectories to define object-centric rewards. These rewards are used to train a reinforcement learning (RL) policy in simulation. The learned policy is then distilled into an image-conditioned diffusion policy using synthetic rollouts rendered with varied viewpoints and lighting. To transfer to the real world, X-Sim introduces an online domain adaptation technique that aligns real and simulated observations during deployment. Importantly, X-Sim does not require any robot teleoperation data. We evaluate it across 5 manipulation tasks in 2 environments and show that it: (1) improves task progress by 30% on average over hand-tracking and sim-to-real baselines, (2) matches behavior cloning with 10x less data collection time, and (3) generalizes to new camera viewpoints and test-time changes. Code and videos are available at https://portal-cornell.github.io/X-Sim/.
Generalist imitation learning policies trained on large datasets show great promise for solving diverse manipulation tasks. However, to ensure generalization to different conditions, policies need to be trained with data collected across a large set of environmental factor variations (e.g., camera pose, table height, distractors) $-$ a prohibitively expensive undertaking, if done exhaustively. We introduce a principled method for deciding what data to collect and how much to collect for each factor by constructing factored scaling curves (FSC), which quantify how policy performance varies as data scales along individual or paired factors. These curves enable targeted data acquisition for the most influential factor combinations within a given budget. We evaluate the proposed method through extensive simulated and real-world experiments, across both training-from-scratch and fine-tuning settings, and show that it boosts success rates in real-world tasks in new environments by up to 26% over existing data-collection strategies. We further demonstrate how factored scaling curves can effectively guide data collection using an offline metric, without requiring real-world evaluation at scale.
Next Best View (NBV) algorithms aim to acquire an optimal set of images using minimal resources, time, or number of captures to enable efficient 3D reconstruction of a scene. Existing approaches often rely on prior scene knowledge or additional image captures and often develop policies that maximize coverage. Yet, for many real scenes with complex geometry and self-occlusions, coverage maximization does not lead to better reconstruction quality directly. In this paper, we propose the View Introspection Network (VIN), which is trained to predict the reconstruction quality improvement of views directly, and the VIN-NBV policy. A greedy sequential sampling-based policy, where at each acquisition step, we sample multiple query views and choose the one with the highest VIN predicted improvement score. We design the VIN to perform 3D-aware featurization of the reconstruction built from prior acquisitions, and for each query view create a feature that can be decoded into an improvement score. We then train the VIN using imitation learning to predict the reconstruction improvement score. We show that VIN-NBV improves reconstruction quality by ~30% over a coverage maximization baseline when operating with constraints on the number of acquisitions or the time in motion.
High-frequency trading (HFT) is an investing strategy that continuously monitors market states and places bid and ask orders at millisecond speeds. Traditional HFT approaches fit models with historical data and assume that future market states follow similar patterns. This limits the effectiveness of any single model to the specific conditions it was trained for. Additionally, these models achieve optimal solutions only under specific market conditions, such as assumptions about stock price's stochastic process, stable order flow, and the absence of sudden volatility. Real-world markets, however, are dynamic, diverse, and frequently volatile. To address these challenges, we propose the FlowHFT, a novel imitation learning framework based on flow matching policy. FlowHFT simultaneously learns strategies from numerous expert models, each proficient in particular market scenarios. As a result, our framework can adaptively adjust investment decisions according to the prevailing market state. Furthermore, FlowHFT incorporates a grid-search fine-tuning mechanism. This allows it to refine strategies and achieve superior performance even in complex or extreme market scenarios where expert strategies may be suboptimal. We test FlowHFT in multiple market environments. We first show that flow matching policy is applicable in stochastic market environments, thus enabling FlowHFT to learn trading strategies under different market conditions. Notably, our single framework consistently achieves performance superior to the best expert for each market condition.
This paper introduces a novel approach to contextual stochastic optimization, integrating operations research and machine learning to address decision-making under uncertainty. Traditional methods often fail to leverage contextual information, which underscores the necessity for new algorithms. In this study, we utilize neural networks with combinatorial optimization layers to encode policies. Our goal is to minimize the empirical risk, which is estimated from past data on uncertain parameters and contexts. To that end, we present a surrogate learning problem and a generic primal-dual algorithm that is applicable to various combinatorial settings in stochastic optimization. Our approach extends classic Fenchel-Young loss results and introduces a new regularization method using sparse perturbations on the distribution simplex. This allows for tractable updates in the original space and can accommodate diverse objective functions. We demonstrate the linear convergence of our algorithm under certain conditions and provide a bound on the non-optimality of the resulting policy in terms of the empirical risk. Experiments on a contextual stochastic minimum weight spanning tree problem show that our algorithm is efficient and scalable, achieving performance comparable to imitation learning of solutions computed using an expensive Lagrangian-based heuristic.
Learning robot policies using imitation learning requires collecting large amounts of costly action-labeled expert demonstrations, which fundamentally limits the scale of training data. A promising approach to address this bottleneck is to harness the abundance of unlabeled observations-e.g., from video demonstrations-to learn latent action labels in an unsupervised way. However, we find that existing methods struggle when applied to complex robot tasks requiring fine-grained motions. We design continuous latent action models (CLAM) which incorporate two key ingredients we find necessary for learning to solve complex continuous control tasks from unlabeled observation data: (a) using continuous latent action labels instead of discrete representations, and (b) jointly training an action decoder to ensure that the latent action space can be easily grounded to real actions with relatively few labeled examples. Importantly, the labeled examples can be collected from non-optimal play data, enabling CLAM to learn performant policies without access to any action-labeled expert data. We demonstrate on continuous control benchmarks in DMControl (locomotion) and MetaWorld (manipulation), as well as on a real WidowX robot arm that CLAM significantly outperforms prior state-of-the-art methods, remarkably with a 2-3x improvement in task success rate compared to the best baseline. Videos and code can be found at clamrobot.github.io.
Learning bimanual manipulation is challenging due to its high dimensionality and tight coordination required between two arms. Eye-in-hand imitation learning, which uses wrist-mounted cameras, simplifies perception by focusing on task-relevant views. However, collecting diverse demonstrations remains costly, motivating the need for scalable data augmentation. While prior work has explored visual augmentation in single-arm settings, extending these approaches to bimanual manipulation requires generating viewpoint-consistent observations across both arms and producing corresponding action labels that are both valid and feasible. In this work, we propose Diffusion for COordinated Dual-arm Data Augmentation (D-CODA), a method for offline data augmentation tailored to eye-in-hand bimanual imitation learning that trains a diffusion model to synthesize novel, viewpoint-consistent wrist-camera images for both arms while simultaneously generating joint-space action labels. It employs constrained optimization to ensure that augmented states involving gripper-to-object contacts adhere to constraints suitable for bimanual coordination. We evaluate D-CODA on 5 simulated and 3 real-world tasks. Our results across 2250 simulation trials and 300 real-world trials demonstrate that it outperforms baselines and ablations, showing its potential for scalable data augmentation in eye-in-hand bimanual manipulation. Our project website is at: https://dcodaaug.github.io/D-CODA/.
Real-world robot manipulation in dynamic unstructured environments requires lifelong adaptability to evolving objects, scenes and tasks. Traditional imitation learning relies on static training paradigms, which are ill-suited for lifelong adaptation. Although Continual Imitation Learnin (CIL) enables incremental task adaptation while preserving learned knowledge, current CIL methods primarily overlook the intrinsic skill characteristics of robot manipulation or depend on manually defined and rigid skills, leading to suboptimal cross-task knowledge transfer. To address these issues, we propose Skill Prompts-based HiErarchical Continual Imitation Learning (SPECI), a novel end-to-end hierarchical CIL policy architecture for robot manipulation. The SPECI framework consists of a multimodal perception and fusion module for heterogeneous sensory information encoding, a high-level skill inference module for dynamic skill extraction and selection, and a low-level action execution module for precise action generation. To enable efficient knowledge transfer on both skill and task levels, SPECI performs continual implicit skill acquisition and reuse via an expandable skill codebook and an attention-driven skill selection mechanism. Furthermore, we introduce mode approximation to augment the last two modules with task-specific and task-sharing parameters, thereby enhancing task-level knowledge transfer. Extensive experiments on diverse manipulation task suites demonstrate that SPECI consistently outperforms state-of-the-art CIL methods across all evaluated metrics, revealing exceptional bidirectional knowledge transfer and superior overall performance.
Today's robots learn new tasks by imitating human examples. However, this standard approach to visual imitation learning is fundamentally limited: the robot observes what the human does, but not why the human chooses those behaviors. Without understanding the features that factor into the human's decisions, robot learners often misinterpret the data and fail to perform the task when the environment changes. We therefore propose a shift in perspective: instead of asking human teachers just to show what actions the robot should take, we also enable humans to indicate task-relevant features using markers and language prompts. Our proposed algorithm, CIVIL, leverages this augmented data to filter the robot's visual observations and extract a feature representation that causally informs human actions. CIVIL then applies these causal features to train a transformer-based policy that emulates human behaviors without being confused by visual distractors. Our simulations, real-world experiments, and user study demonstrate that robots trained with CIVIL can learn from fewer human demonstrations and perform better than state-of-the-art baselines, especially in previously unseen scenarios. See videos at our project website: https://civil2025.github.io
Learning from few demonstrations to develop policies robust to variations in robot initial positions and object poses is a problem of significant practical interest in robotics. Compared to imitation learning, which often struggles to generalize from limited samples, reinforcement learning (RL) can autonomously explore to obtain robust behaviors. Training RL agents through direct interaction with the real world is often impractical and unsafe, while building simulation environments requires extensive manual effort, such as designing scenes and crafting task-specific reward functions. To address these challenges, we propose an integrated real-to-sim-to-real pipeline that constructs simulation environments based on expert demonstrations by identifying scene objects from images and retrieving their corresponding 3D models from existing libraries. We introduce a projection-based reward model for RL policy training that is supervised by a vision-language model (VLM) using human-guided object projection relationships as prompts, with the policy further fine-tuned using expert demonstrations. In general, our work focuses on the construction of simulation environments and RL-based policy training, ultimately enabling the deployment of reliable robotic control policies in real-world scenarios.
This work focuses on the design of a deep learning-based autonomous driving system deployed and tested on the real-world MIT Racecar to assess its effectiveness in driving scenarios. The Deep Neural Network (DNN) translates raw image inputs into real-time steering commands in an end-to-end learning fashion, following the imitation learning framework. The key design challenge is to ensure that DNN predictions are accurate and fast enough, at a high sampling frequency, and result in smooth vehicle operation under different operating conditions. In this study, we design and compare various DNNs, to identify the most effective approach for real-time autonomous driving. In designing the DNNs, we adopted an incremental design approach that involved enhancing the model capacity and dataset to address the challenges of real-world driving scenarios. We designed a PD system, CNN, CNN-LSTM, and CNN-NODE, and evaluated their performance on the real-world MIT Racecar. While the PD system handled basic lane following, it struggled with sharp turns and lighting variations. The CNN improved steering but lacked temporal awareness, which the CNN-LSTM addressed as it resulted in smooth driving performance. The CNN-NODE performed similarly to the CNN-LSTM in handling driving dynamics, yet with slightly better driving performance. The findings of this research highlight the importance of iterative design processes in developing robust DNNs for autonomous driving applications. The experimental video is available at https://www.youtube.com/watch?v=FNNYgU--iaY.
Imitation learning holds the promise of equipping robots with versatile skills by learning from expert demonstrations. However, policies trained on finite datasets often struggle to generalize beyond the training distribution. In this work, we present a unified perspective on the generalization capability of imitation learning, grounded in both information theorey and data distribution property. We first show that the generalization gap can be upper bounded by (i) the conditional information bottleneck on intermediate representations and (ii) the mutual information between the model parameters and the training dataset. This characterization provides theoretical guidance for designing effective training strategies in imitation learning, particularly in determining whether to freeze, fine-tune, or train large pretrained encoders (e.g., vision-language models or vision foundation models) from scratch to achieve better generalization. Furthermore, we demonstrate that high conditional entropy from input to output induces a flatter likelihood landscape, thereby reducing the upper bound on the generalization gap. In addition, it shortens the stochastic gradient descent (SGD) escape time from sharp local minima, which may increase the likelihood of reaching global optima under fixed optimization budgets. These insights explain why imitation learning often exhibits limited generalization and underscore the importance of not only scaling the diversity of input data but also enriching the variability of output labels conditioned on the same input.
Imitation Learning (IL) techniques aim to replicate human behaviors in specific tasks. While IL has gained prominence due to its effectiveness and efficiency, traditional methods often focus on datasets collected from experts to produce a single efficient policy. Recently, extensions have been proposed to handle datasets of diverse behaviors by mainly focusing on learning transition-level diverse policies or on performing entropy maximization at the trajectory level. While these methods may lead to diverse behaviors, they may not be sufficient to reproduce the actual diversity of demonstrations or to allow controlled trajectory generation. To overcome these drawbacks, we propose a different method based on two key features: a) Temporal Consistency that ensures consistent behaviors across entire episodes and not just at the transition level as well as b) Controllability obtained by constructing a latent space of behaviors that allows users to selectively activate specific behaviors based on their requirements. We compare our approach to state-of-the-art methods over a diverse set of tasks and environments. Project page: https://mathieu-petitbois.github.io/projects/swr/
Collaboration is ubiquitous and essential in day-to-day life -- from exchanging ideas, to delegating tasks, to generating plans together. This work studies how LLMs can adaptively collaborate to perform complex embodied reasoning tasks. To this end we introduce MINDcraft, an easily extensible platform built to enable LLM agents to control characters in the open-world game of Minecraft; and MineCollab, a benchmark to test the different dimensions of embodied and collaborative reasoning. An experimental study finds that the primary bottleneck in collaborating effectively for current state-of-the-art agents is efficient natural language communication, with agent performance dropping as much as 15% when they are required to communicate detailed task completion plans. We conclude that existing LLM agents are ill-optimized for multi-agent collaboration, especially in embodied scenarios, and highlight the need to employ methods beyond in-context and imitation learning. Our website can be found here: https://mindcraft-minecollab.github.io/
Achieving both realism and controllability in interactive closed-loop traffic simulation remains a key challenge in autonomous driving. Data-driven simulation methods reproduce realistic trajectories but suffer from covariate shift in closed-loop deployment, compounded by simplified dynamics models that further reduce reliability. Conversely, physics-based simulation methods enhance reliable and controllable closed-loop interactions but often lack expert demonstrations, compromising realism. To address these challenges, we introduce a dual-stage AV-centered simulation framework that conducts open-loop imitation learning pre-training in a data-driven simulator to capture trajectory-level realism and multimodality, followed by closed-loop reinforcement learning fine-tuning in a physics-based simulator to enhance controllability and mitigate covariate shift. In the fine-tuning stage, we propose RIFT, a simple yet effective closed-loop RL fine-tuning strategy that preserves the trajectory-level multimodality through a GRPO-style group-relative advantage formulation, while enhancing controllability and training stability by replacing KL regularization with the dual-clip mechanism. Extensive experiments demonstrate that RIFT significantly improves the realism and controllability of generated traffic scenarios, providing a robust platform for evaluating autonomous vehicle performance in diverse and interactive scenarios.