Autonomous agents capable of diverse object manipulations should be able to acquire a wide range of manipulation skills with high reusability. Although advances in deep learning have made it increasingly feasible to replicate the dexterity of human teleoperation in robots, generalizing these acquired skills to previously unseen scenarios remains a significant challenge. In this study, we propose a novel algorithm, Gaze-based Bottleneck-aware Robot Manipulation (GazeBot), which enables high reusability of the learned motions even when the object positions and end-effector poses differ from those in the provided demonstrations. By leveraging gaze information and motion bottlenecks, both crucial features for object manipulation, GazeBot achieves high generalization performance compared with state-of-the-art imitation learning methods, without sacrificing its dexterity and reactivity. Furthermore, the training process of GazeBot is entirely data-driven once a demonstration dataset with gaze data is provided. Videos and code are available at https://crumbyrobotics.github.io/gazebot.
Semantic communication marks a new paradigm shift from bit-wise data transmission to semantic information delivery for the purpose of bandwidth reduction. To more effectively carry out specialized downstream tasks at the receiver end, it is crucial to define the most critical semantic message in the data based on the task or goal-oriented features. In this work, we propose a novel goal-oriented communication (GO-COM) framework, namely Goal-Oriented Semantic Variational Autoencoder (GOS-VAE), by focusing on the extraction of the semantics vital to the downstream tasks. Specifically, we adopt a Vector Quantized Variational Autoencoder (VQ-VAE) to compress media data at the transmitter side. Instead of targeting the pixel-wise image data reconstruction, we measure the quality-of-service at the receiver end based on a pre-defined task-incentivized model. Moreover, to capture the relevant semantic features in the data reconstruction, imitation learning is adopted to measure the data regeneration quality in terms of goal-oriented semantics. Our experimental results demonstrate the power of imitation learning in characterizing goal-oriented semantics and bandwidth efficiency of our proposed GOS-VAE.
Successful collaboration requires team members to stay aligned, especially in complex sequential tasks. Team members must dynamically coordinate which subtasks to perform and in what order. However, real-world constraints like partial observability and limited communication bandwidth often lead to suboptimal collaboration. Even among expert teams, the same task can be executed in multiple ways. To develop multi-agent systems and human-AI teams for such tasks, we are interested in data-driven learning of multimodal team behaviors. Multi-Agent Imitation Learning (MAIL) provides a promising framework for data-driven learning of team behavior from demonstrations, but existing methods struggle with heterogeneous demonstrations, as they assume that all demonstrations originate from a single team policy. Hence, in this work, we introduce DTIL: a hierarchical MAIL algorithm designed to learn multimodal team behaviors in complex sequential tasks. DTIL represents each team member with a hierarchical policy and learns these policies from heterogeneous team demonstrations in a factored manner. By employing a distribution-matching approach, DTIL mitigates compounding errors and scales effectively to long horizons and continuous state representations. Experimental results show that DTIL outperforms MAIL baselines and accurately models team behavior across a variety of collaborative scenarios.
Reinforcement learning plays a crucial role in vehicle control by guiding agents to learn optimal control strategies through designing or learning appropriate reward signals. However, in vehicle control applications, rewards typically need to be manually designed while considering multiple implicit factors, which easily introduces human biases. Although imitation learning methods does not rely on explicit reward signals, they necessitate high-quality expert actions, which are often challenging to acquire. To address these issues, we propose a reward-free reinforcement learning framework (RFRLF). This framework directly learns the target states to optimize agent behavior through a target state prediction network (TSPN) and a reward-free state-guided policy network (RFSGPN), avoiding the dependence on manually designed reward signals. Specifically, the policy network is learned via minimizing the differences between the predicted state and the expert state. Experimental results demonstrate the effectiveness of the proposed RFRLF in controlling vehicle driving, showing its advantages in improving learning efficiency and adapting to reward-free environments.
Robotics has long sought to develop visual-servoing robots capable of completing previously unseen long-horizon tasks. Hierarchical approaches offer a pathway for achieving this goal by executing skill combinations arranged by a task planner, with each visuomotor skill pre-trained using a specific imitation learning (IL) algorithm. However, even in simple long-horizon tasks like skill chaining, hierarchical approaches often struggle due to a problem we identify as Observation Space Shift (OSS), where the sequential execution of preceding skills causes shifts in the observation space, disrupting the performance of subsequent individually trained skill policies. To validate OSS and evaluate its impact on long-horizon tasks, we introduce BOSS (a Benchmark for Observation Space Shift). BOSS comprises three distinct challenges: "Single Predicate Shift", "Accumulated Predicate Shift", and "Skill Chaining", each designed to assess a different aspect of OSS's negative effect. We evaluated several recent popular IL algorithms on BOSS, including three Behavioral Cloning methods and the Visual Language Action model OpenVLA. Even on the simplest challenge, we observed average performance drops of 67%, 35%, 34%, and 54%, respectively, when comparing skill performance with and without OSS. Additionally, we investigate a potential solution to OSS that scales up the training data for each skill with a larger and more visually diverse set of demonstrations, with our results showing it is not sufficient to resolve OSS. The project page is: https://boss-benchmark.github.io/
We explore the potential of large-scale generative video models for autonomous driving, introducing an open-source auto-regressive video model (VaViM) and its companion video-action model (VaVAM) to investigate how video pre-training transfers to real-world driving. VaViM is a simple auto-regressive video model that predicts frames using spatio-temporal token sequences. We show that it captures the semantics and dynamics of driving scenes. VaVAM, the video-action model, leverages the learned representations of VaViM to generate driving trajectories through imitation learning. Together, the models form a complete perception-to-action pipeline. We evaluate our models in open- and closed-loop driving scenarios, revealing that video-based pre-training holds promise for autonomous driving. Key insights include the semantic richness of the learned representations, the benefits of scaling for video synthesis, and the complex relationship between model size, data, and safety metrics in closed-loop evaluations. We release code and model weights at https://github.com/valeoai/VideoActionModel
The development of a generalist agent capable of solving a wide range of sequential decision-making tasks remains a significant challenge. We address this problem in a cross-agent setup where agents share the same observation space but differ in their action spaces. Our approach builds on the universal policy framework, which decouples policy learning into two stages: a diffusion-based planner that generates observation sequences and an inverse dynamics model that assigns actions to these plans. We propose a method for training the planner on a joint dataset composed of trajectories from all agents. This method offers the benefit of positive transfer by pooling data from different agents, while the primary challenge lies in adapting shared plans to each agent's unique constraints. We evaluate our approach on the BabyAI environment, covering tasks of varying complexity, and demonstrate positive transfer across agents. Additionally, we examine the planner's generalisation ability to unseen agents and compare our method to traditional imitation learning approaches. By training on a pooled dataset from multiple agents, our universal policy achieves an improvement of up to $42.20\%$ in task completion accuracy compared to a policy trained on a dataset from a single agent.
Human motion is highly diverse and dynamic, posing challenges for imitation learning algorithms that aim to generalize motor skills for controlling simulated characters. Previous methods typically rely on a universal full-body controller for tracking reference motion (tracking-based model) or a unified full-body skill embedding space (skill embedding). However, these approaches often struggle to generalize and scale to larger motion datasets. In this work, we introduce a novel skill learning framework, ModSkill, that decouples complex full-body skills into compositional, modular skills for independent body parts. Our framework features a skill modularization attention layer that processes policy observations into modular skill embeddings that guide low-level controllers for each body part. We also propose an Active Skill Learning approach with Generative Adaptive Sampling, using large motion generation models to adaptively enhance policy learning in challenging tracking scenarios. Our results show that this modularized skill learning framework, enhanced by generative sampling, outperforms existing methods in precise full-body motion tracking and enables reusable skill embeddings for diverse goal-driven tasks.
Automated Planning algorithms require a model of the domain that specifies the preconditions and effects of each action. Obtaining such a domain model is notoriously hard. Algorithms for learning domain models exist, yet it remains unclear whether learning a domain model and planning is an effective approach for numeric planning environments, i.e., where states include discrete and numeric state variables. In this work, we explore the benefits of learning a numeric domain model and compare it with alternative model-free solutions. As a case study, we use two tasks in Minecraft, a popular sandbox game that has been used as an AI challenge. First, we consider an offline learning setting, where a set of expert trajectories are available to learn from. This is the standard setting for learning domain models. We used the Numeric Safe Action Model Learning (NSAM) algorithm to learn a numeric domain model and solve new problems with the learned domain model and a numeric planner. We call this model-based solution NSAM_(+p), and compare it to several model-free Imitation Learning (IL) and Offline Reinforcement Learning (RL) algorithms. Empirical results show that some IL algorithms can learn faster to solve simple tasks, while NSAM_(+p) allows solving tasks that require long-term planning and enables generalizing to solve problems in larger environments. Then, we consider an online learning setting, where learning is done by moving an agent in the environment. For this setting, we introduce RAMP. In RAMP, observations collected during the agent's execution are used to simultaneously train an RL policy and learn a planning domain action model. This forms a positive feedback loop between the RL policy and the learned domain model. We demonstrate experimentally the benefits of using RAMP, showing that it finds more efficient plans and solves more problems than several RL baselines.
Diffusion policies have shown promise in learning complex behaviors from demonstrations, particularly for tasks requiring precise control and long-term planning. However, they face challenges in robustness when encountering distribution shifts. This paper explores improving diffusion-based imitation learning models through online interactions with the environment. We propose OTPR (Optimal Transport-guided score-based diffusion Policy for Reinforcement learning fine-tuning), a novel method that integrates diffusion policies with RL using optimal transport theory. OTPR leverages the Q-function as a transport cost and views the policy as an optimal transport map, enabling efficient and stable fine-tuning. Moreover, we introduce masked optimal transport to guide state-action matching using expert keypoints and a compatibility-based resampling strategy to enhance training stability. Experiments on three simulation tasks demonstrate OTPR's superior performance and robustness compared to existing methods, especially in complex and sparse-reward environments. In sum, OTPR provides an effective framework for combining IL and RL, achieving versatile and reliable policy learning. The code will be released at https://github.com/Sunmmyy/OTPR.git.
We study the problem of reinforcement learning in infinite-horizon discounted linear Markov decision processes (MDPs), and propose the first computationally efficient algorithm achieving near-optimal regret guarantees in this setting. Our main idea is to combine two classic techniques for optimistic exploration: additive exploration bonuses applied to the reward function, and artificial transitions made to an absorbing state with maximal return. We show that, combined with a regularized approximate dynamic-programming scheme, the resulting algorithm achieves a regret of order $\tilde{\mathcal{O}} (\sqrt{d^3 (1 - \gamma)^{- 7 / 2} T})$, where $T$ is the total number of sample transitions, $\gamma \in (0,1)$ is the discount factor, and $d$ is the feature dimensionality. The results continue to hold against adversarial reward sequences, enabling application of our method to the problem of imitation learning in linear MDPs, where we achieve state-of-the-art results.
Many everyday mobile manipulation tasks require precise interaction with small objects, such as grasping a knob to open a cabinet or pressing a light switch. In this paper, we develop Servoing with Vision Models (SVM), a closed-loop training-free framework that enables a mobile manipulator to tackle such precise tasks involving the manipulation of small objects. SVM employs an RGB-D wrist camera and uses visual servoing for control. Our novelty lies in the use of state-of-the-art vision models to reliably compute 3D targets from the wrist image for diverse tasks and under occlusion due to the end-effector. To mitigate occlusion artifacts, we employ vision models to out-paint the end-effector thereby significantly enhancing target localization. We demonstrate that aided by out-painting methods, open-vocabulary object detectors can serve as a drop-in module to identify semantic targets (e.g. knobs) and point tracking methods can reliably track interaction sites indicated by user clicks. This training-free method obtains an 85% zero-shot success rate on manipulating unseen objects in novel environments in the real world, outperforming an open-loop control method and an imitation learning baseline trained on 1000+ demonstrations by an absolute success rate of 50%.
Imitation learning techniques have been shown to be highly effective in real-world control scenarios, such as robotics. However, these approaches not only suffer from compounding error issues but also require human experts to provide complete trajectories. Although there exist interactive methods where an expert oversees the robot and intervenes if needed, these extensions usually only utilize the data collected during intervention periods and ignore the feedback signal hidden in non-intervention timesteps. In this work, we create a model to formulate how the interventions occur in such cases, and show that it is possible to learn a policy with just a handful of expert interventions. Our key insight is that it is possible to get crucial information about the quality of the current state and the optimality of the chosen action from expert feedback, regardless of the presence or the absence of intervention. We evaluate our method on various discrete and continuous simulation environments, a real-world robotic manipulation task, as well as a human subject study. Videos and the code can be found at https://liralab.usc.edu/mile .
Existing end-to-end autonomous driving (AD) algorithms typically follow the Imitation Learning (IL) paradigm, which faces challenges such as causal confusion and the open-loop gap. In this work, we establish a 3DGS-based closed-loop Reinforcement Learning (RL) training paradigm. By leveraging 3DGS techniques, we construct a photorealistic digital replica of the real physical world, enabling the AD policy to extensively explore the state space and learn to handle out-of-distribution scenarios through large-scale trial and error. To enhance safety, we design specialized rewards that guide the policy to effectively respond to safety-critical events and understand real-world causal relationships. For better alignment with human driving behavior, IL is incorporated into RL training as a regularization term. We introduce a closed-loop evaluation benchmark consisting of diverse, previously unseen 3DGS environments. Compared to IL-based methods, RAD achieves stronger performance in most closed-loop metrics, especially 3x lower collision rate. Abundant closed-loop results are presented at https://hgao-cv.github.io/RAD.
Recent advances in imitation learning, particularly using generative modelling techniques like diffusion, have enabled policies to capture complex multi-modal action distributions. However, these methods often require large datasets and multiple inference steps for action generation, posing challenges in robotics where the cost for data collection is high and computation resources are limited. To address this, we introduce IMLE Policy, a novel behaviour cloning approach based on Implicit Maximum Likelihood Estimation (IMLE). IMLE Policy excels in low-data regimes, effectively learning from minimal demonstrations and requiring 38\% less data on average to match the performance of baseline methods in learning complex multi-modal behaviours. Its simple generator-based architecture enables single-step action generation, improving inference speed by 97.3\% compared to Diffusion Policy, while outperforming single-step Flow Matching. We validate our approach across diverse manipulation tasks in simulated and real-world environments, showcasing its ability to capture complex behaviours under data constraints. Videos and code are provided on our project page: https://imle-policy.github.io/.
Learning for manipulation requires using policies that have access to rich sensory information such as point clouds or RGB images. Point clouds efficiently capture geometric structures, making them essential for manipulation tasks in imitation learning. In contrast, RGB images provide rich texture and semantic information that can be crucial for certain tasks. Existing approaches for fusing both modalities assign 2D image features to point clouds. However, such approaches often lose global contextual information from the original images. In this work, we propose FPV-Net, a novel imitation learning method that effectively combines the strengths of both point cloud and RGB modalities. Our method conditions the point-cloud encoder on global and local image tokens using adaptive layer norm conditioning, leveraging the beneficial properties of both modalities. Through extensive experiments on the challenging RoboCasa benchmark, we demonstrate the limitations of relying on either modality alone and show that our method achieves state-of-the-art performance across all tasks.
Generalizable humanoid loco-manipulation poses significant challenges, requiring coordinated whole-body control and precise, contact-rich object manipulation. To address this, this paper introduces HOMIE, a semi-autonomous teleoperation system that combines a reinforcement learning policy for body control mapped to a pedal, an isomorphic exoskeleton arm for arm control, and motion-sensing gloves for hand control, forming a unified cockpit to freely operate humanoids and establish a data flywheel. The policy incorporates novel designs, including an upper-body pose curriculum, a height-tracking reward, and symmetry utilization. These features enable the system to perform walking and squatting to specific heights while seamlessly adapting to arbitrary upper-body poses. The exoskeleton, by eliminating the reliance on inverse dynamics, delivers faster and more precise arm control. The gloves utilize Hall sensors instead of servos, allowing even compact devices to achieve 15 or more degrees of freedom and freely adapt to any model of dexterous hands. Compared to previous teleoperation systems, HOMIE stands out for its exceptional efficiency, completing tasks in half the time; its expanded working range, allowing users to freely reach high and low areas as well as interact with any objects; and its affordability, with a price of just $500. The system is fully open-source, demos and code can be found in our https://homietele.github.io/.
Next-token prediction with the logarithmic loss is a cornerstone of autoregressive sequence modeling, but, in practice, suffers from error amplification, where errors in the model compound and generation quality degrades as sequence length $H$ increases. From a theoretical perspective, this phenomenon should not appear in well-specified settings, and, indeed, a growing body of empirical work hypothesizes that misspecification, where the learner is not sufficiently expressive to represent the target distribution, may be the root cause. Under misspecification -- where the goal is to learn as well as the best-in-class model up to a multiplicative approximation factor $C\geq 1$ -- we confirm that $C$ indeed grows with $H$ for next-token prediction, lending theoretical support to this empirical hypothesis. We then ask whether this mode of error amplification is avoidable algorithmically, computationally, or information-theoretically, and uncover inherent computational-statistical tradeoffs. We show: (1) Information-theoretically, one can avoid error amplification and achieve $C=O(1)$. (2) Next-token prediction can be made robust so as to achieve $C=\tilde O(H)$, representing moderate error amplification, but this is an inherent barrier: any next-token prediction-style objective must suffer $C=\Omega(H)$. (3) For the natural testbed of autoregressive linear models, no computationally efficient algorithm can achieve sub-polynomial approximation factor $C=e^{(\log H)^{1-\Omega(1)}}$; however, at least for binary token spaces, one can smoothly trade compute for statistical power and improve on $C=\Omega(H)$ in sub-exponential time. Our results have consequences in the more general setting of imitation learning, where the widely-used behavior cloning algorithm generalizes next-token prediction.
Designing modern imitation learning (IL) policies requires making numerous decisions, including the selection of feature encoding, architecture, policy representation, and more. As the field rapidly advances, the range of available options continues to grow, creating a vast and largely unexplored design space for IL policies. In this work, we present X-IL, an accessible open-source framework designed to systematically explore this design space. The framework's modular design enables seamless swapping of policy components, such as backbones (e.g., Transformer, Mamba, xLSTM) and policy optimization techniques (e.g., Score-matching, Flow-matching). This flexibility facilitates comprehensive experimentation and has led to the discovery of novel policy configurations that outperform existing methods on recent robot learning benchmarks. Our experiments demonstrate not only significant performance gains but also provide valuable insights into the strengths and weaknesses of various design choices. This study serves as both a practical reference for practitioners and a foundation for guiding future research in imitation learning.
Although Reinforcement Learning (RL)-based Traffic Signal Control (TSC) methods have been extensively studied, their practical applications still raise some serious issues such as high learning cost and poor generalizability. This is because the ``trial-and-error'' training style makes RL agents extremely dependent on the specific traffic environment, which also requires a long convergence time. To address these issues, we propose a novel Federated Imitation Learning (FIL)-based framework for multi-intersection TSC, named FitLight, which allows RL agents to plug-and-play for any traffic environment without additional pre-training cost. Unlike existing imitation learning approaches that rely on pre-training RL agents with demonstrations, FitLight allows real-time imitation learning and seamless transition to reinforcement learning. Due to our proposed knowledge-sharing mechanism and novel hybrid pressure-based agent design, RL agents can quickly find a best control policy with only a few episodes. Moreover, for resource-constrained TSC scenarios, FitLight supports model pruning and heterogeneous model aggregation, such that RL agents can work on a micro-controller with merely 16{\it KB} RAM and 32{\it KB} ROM. Extensive experiments demonstrate that, compared to state-of-the-art methods, FitLight not only provides a superior starting point but also converges to a better final solution on both real-world and synthetic datasets, even under extreme resource limitations.