Dexterous manipulation is a crucial yet highly complex challenge in humanoid robotics, demanding precise, adaptable, and sample-efficient learning methods. As humanoid robots are usually designed to operate in human-centric environments and interact with everyday objects, mastering dexterous manipulation is critical for real-world deployment. Traditional approaches, such as reinforcement learning and imitation learning, have made significant strides, but they often struggle due to the unique challenges of real-world dexterous manipulation, including high-dimensional control, limited training data, and covariate shift. This survey provides a comprehensive overview of these challenges and reviews existing learning-based methods for dexterous manipulation, spanning imitation learning, reinforcement learning, and hybrid approaches. A promising yet underexplored direction is interactive imitation learning, where human feedback actively refines a robot's behavior during training. While interactive imitation learning has shown success in various robotic tasks, its application to dexterous manipulation remains limited. To address this gap, we examine current interactive imitation learning techniques applied to other robotic tasks and discuss how these methods can be adapted to enhance dexterous manipulation. By synthesizing state-of-the-art research, this paper highlights key challenges, identifies gaps in current methodologies, and outlines potential directions for leveraging interactive imitation learning to improve dexterous robotic skills.
The 3D scene graph models spatial relationships between objects, enabling the agent to efficiently navigate in a partially observable environment and predict the location of the target object.This paper proposes an original framework named SGN-CIRL (3D Scene Graph-Based Reinforcement Learning Navigation) for mapless reinforcement learning-based robot navigation with learnable representation of open-vocabulary 3D scene graph. To accelerate and stabilize the training of reinforcement learning-based algorithms, the framework also employs imitation learning and curriculum learning. The first one enables the agent to learn from demonstrations, while the second one structures the training process by gradually increasing task complexity from simple to more advanced scenarios. Numerical experiments conducted in the Isaac Sim environment showed that using a 3D scene graph for reinforcement learning significantly increased the success rate in difficult navigation cases. The code is open-sourced and available at: https://github.com/Xisonik/Aloha\_graph.
Autonomous driving promises significant advancements in mobility, road safety and traffic efficiency, yet reinforcement learning and imitation learning face safe-exploration and distribution-shift challenges. Although human-AI collaboration alleviates these issues, it often relies heavily on extensive human intervention, which increases costs and reduces efficiency. This paper develops a confidence-guided human-AI collaboration (C-HAC) strategy to overcome these limitations. First, C-HAC employs a distributional proxy value propagation method within the distributional soft actor-critic (DSAC) framework. By leveraging return distributions to represent human intentions C-HAC achieves rapid and stable learning of human-guided policies with minimal human interaction. Subsequently, a shared control mechanism is activated to integrate the learned human-guided policy with a self-learning policy that maximizes cumulative rewards. This enables the agent to explore independently and continuously enhance its performance beyond human guidance. Finally, a policy confidence evaluation algorithm capitalizes on DSAC's return distribution networks to facilitate dynamic switching between human-guided and self-learning policies via a confidence-based intervention function. This ensures the agent can pursue optimal policies while maintaining safety and performance guarantees. Extensive experiments across diverse driving scenarios reveal that C-HAC significantly outperforms conventional methods in terms of safety, efficiency, and overall performance, achieving state-of-the-art results. The effectiveness of the proposed method is further validated through real-world road tests in complex traffic conditions. The videos and code are available at: https://github.com/lzqw/C-HAC.
Understanding and predicting articulated actions is important in robot learning. However, common architectures such as MLPs and Transformers lack inductive biases that reflect the underlying kinematic structure of articulated systems. To this end, we propose the Neural Rodrigues Operator, a learnable generalization of the classical forward kinematics operation, designed to inject kinematics-aware inductive bias into neural computation. Building on this operator, we design the Rodrigues Network (RodriNet), a novel neural architecture specialized for processing actions. We evaluate the expressivity of our network on two synthetic tasks on kinematic and motion prediction, showing significant improvements compared to standard backbones. We further demonstrate its effectiveness in two realistic applications: (i) imitation learning on robotic benchmarks with the Diffusion Policy, and (ii) single-image 3D hand reconstruction. Our results suggest that integrating structured kinematic priors into the network architecture improves action learning in various domains.
Language-instructed active object localization is a critical challenge for robots, requiring efficient exploration of partially observable environments. However, state-of-the-art approaches either struggle to generalize beyond demonstration datasets (e.g., imitation learning methods) or fail to generate physically grounded actions (e.g., VLMs). To address these limitations, we introduce WoMAP (World Models for Active Perception): a recipe for training open-vocabulary object localization policies that: (i) uses a Gaussian Splatting-based real-to-sim-to-real pipeline for scalable data generation without the need for expert demonstrations, (ii) distills dense rewards signals from open-vocabulary object detectors, and (iii) leverages a latent world model for dynamics and rewards prediction to ground high-level action proposals at inference time. Rigorous simulation and hardware experiments demonstrate WoMAP's superior performance in a broad range of zero-shot object localization tasks, with more than 9x and 2x higher success rates compared to VLM and diffusion policy baselines, respectively. Further, we show that WoMAP achieves strong generalization and sim-to-real transfer on a TidyBot.
This paper proposes a novel stable learning theory for recurrent neural networks (RNNs), so-called variational adaptive noise and dropout (VAND). As stabilizing factors for RNNs, noise and dropout on the internal state of RNNs have been separately confirmed in previous studies. We reinterpret the optimization problem of RNNs as variational inference, showing that noise and dropout can be derived simultaneously by transforming the explicit regularization term arising in the optimization problem into implicit regularization. Their scale and ratio can also be adjusted appropriately to optimize the main objective of RNNs, respectively. In an imitation learning scenario with a mobile manipulator, only VAND is able to imitate sequential and periodic behaviors as instructed. https://youtu.be/UOho3Xr6A2w
Modern reinforcement learning (RL) algorithms have found success by using powerful probabilistic models, such as transformers, energy-based models, and diffusion/flow-based models. To this end, RL researchers often choose to pay the price of accommodating these models into their algorithms -- diffusion models are expressive, but are computationally intensive due to their reliance on solving differential equations, while autoregressive transformer models are scalable but typically require learning discrete representations. Normalizing flows (NFs), by contrast, seem to provide an appealing alternative, as they enable likelihoods and sampling without solving differential equations or autoregressive architectures. However, their potential in RL has received limited attention, partly due to the prevailing belief that normalizing flows lack sufficient expressivity. We show that this is not the case. Building on recent work in NFs, we propose a single NF architecture which integrates seamlessly into RL algorithms, serving as a policy, Q-function, and occupancy measure. Our approach leads to much simpler algorithms, and achieves higher performance in imitation learning, offline, goal conditioned RL and unsupervised RL.
Due to their expressive capacity, diffusion models have shown great promise in offline RL and imitation learning. Diffusion Actor-Critic with Entropy Regulator (DACER) extended this capability to online RL by using the reverse diffusion process as a policy approximator, trained end-to-end with policy gradient methods, achieving strong performance. However, this comes at the cost of requiring many diffusion steps, which significantly hampers training efficiency, while directly reducing the steps leads to noticeable performance degradation. Critically, the lack of inference efficiency becomes a significant bottleneck for applying diffusion policies in real-time online RL settings. To improve training and inference efficiency while maintaining or even enhancing performance, we propose a Q-gradient field objective as an auxiliary optimization target to guide the denoising process at each diffusion step. Nonetheless, we observe that the independence of the Q-gradient field from the diffusion time step negatively impacts the performance of the diffusion policy. To address this, we introduce a temporal weighting mechanism that enables the model to efficiently eliminate large-scale noise in the early stages and refine actions in the later stages. Experimental results on MuJoCo benchmarks and several multimodal tasks demonstrate that the DACER2 algorithm achieves state-of-the-art performance in most MuJoCo control tasks with only five diffusion steps, while also exhibiting stronger multimodality compared to DACER.
Imitation Learning has become a fundamental approach in robotic manipulation. However, collecting large-scale real-world robot demonstrations is prohibitively expensive. Simulators offer a cost-effective alternative, but the sim-to-real gap make it extremely challenging to scale. Therefore, we introduce RoboTransfer, a diffusion-based video generation framework for robotic data synthesis. Unlike previous methods, RoboTransfer integrates multi-view geometry with explicit control over scene components, such as background and object attributes. By incorporating cross-view feature interactions and global depth/normal conditions, RoboTransfer ensures geometry consistency across views. This framework allows fine-grained control, including background edits and object swaps. Experiments demonstrate that RoboTransfer is capable of generating multi-view videos with enhanced geometric consistency and visual fidelity. In addition, policies trained on the data generated by RoboTransfer achieve a 33.3% relative improvement in the success rate in the DIFF-OBJ setting and a substantial 251% relative improvement in the more challenging DIFF-ALL scenario. Explore more demos on our project page: https://horizonrobotics.github.io/robot_lab/robotransfer
Learning visual representations from observing actions to benefit robot visuo-motor policy generation is a promising direction that closely resembles human cognitive function and perception. Motivated by this, and further inspired by psychological theories suggesting that humans process scenes in an object-based fashion, we propose an object-centric encoder that performs semantic segmentation and visual representation generation in a coupled manner, unlike other works, which treat these as separate processes. To achieve this, we leverage the Slot Attention mechanism and use the SOLV model, pretrained in large out-of-domain datasets, to bootstrap fine-tuning on human action video data. Through simulated robotic tasks, we demonstrate that visual representations can enhance reinforcement and imitation learning training, highlighting the effectiveness of our integrated approach for semantic segmentation and encoding. Furthermore, we show that exploiting models pretrained on out-of-domain datasets can benefit this process, and that fine-tuning on datasets depicting human actions -- although still out-of-domain -- , can significantly improve performance due to close alignment with robotic tasks. These findings show the capability to reduce reliance on annotated or robot-specific action datasets and the potential to build on existing visual encoders to accelerate training and improve generalizability.
Achieving generalizable and precise robotic manipulation across diverse environments remains a critical challenge, largely due to limitations in spatial perception. While prior imitation-learning approaches have made progress, their reliance on raw RGB inputs and handcrafted features often leads to overfitting and poor 3D reasoning under varied lighting, occlusion, and object conditions. In this paper, we propose a unified framework that couples robust multimodal perception with reliable grasp prediction. Our architecture fuses domain-randomized augmentation, monocular depth estimation, and a depth-aware 6-DoF Grasp Prompt into a single spatial representation for downstream action planning. Conditioned on this encoding and a high-level task prompt, our diffusion-based policy yields precise action sequences, achieving up to 40% improvement in grasp success and 45% higher task success rates under environmental variation. These results demonstrate that spatially grounded perception, paired with diffusion-based imitation learning, offers a scalable and robust solution for general-purpose robotic grasping.
Imitation learning advances robot capabilities by enabling the acquisition of diverse behaviors from human demonstrations. However, large-scale datasets used for policy training often introduce substantial variability in quality, which can negatively impact performance. As a result, automatically curating datasets by filtering low-quality samples to improve quality becomes essential. Existing robotic curation approaches rely on costly manual annotations and perform curation at a coarse granularity, such as the dataset or trajectory level, failing to account for the quality of individual state-action pairs. To address this, we introduce SCIZOR, a self-supervised data curation framework that filters out low-quality state-action pairs to improve the performance of imitation learning policies. SCIZOR targets two complementary sources of low-quality data: suboptimal data, which hinders learning with undesirable actions, and redundant data, which dilutes training with repetitive patterns. SCIZOR leverages a self-supervised task progress predictor for suboptimal data to remove samples lacking task progression, and a deduplication module operating on joint state-action representation for samples with redundant patterns. Empirically, we show that SCIZOR enables imitation learning policies to achieve higher performance with less data, yielding an average improvement of 15.4% across multiple benchmarks. More information is available at: https://ut-austin-rpl.github.io/SCIZOR/
We introduce Behavior from Language and Demonstration (BLADE), a framework for long-horizon robotic manipulation by integrating imitation learning and model-based planning. BLADE leverages language-annotated demonstrations, extracts abstract action knowledge from large language models (LLMs), and constructs a library of structured, high-level action representations. These representations include preconditions and effects grounded in visual perception for each high-level action, along with corresponding controllers implemented as neural network-based policies. BLADE can recover such structured representations automatically, without manually labeled states or symbolic definitions. BLADE shows significant capabilities in generalizing to novel situations, including novel initial states, external state perturbations, and novel goals. We validate the effectiveness of our approach both in simulation and on real robots with a diverse set of objects with articulated parts, partial observability, and geometric constraints.
Vision-language-action (VLA) models have emerged as the next generation of models in robotics. However, despite leveraging powerful pre-trained Vision-Language Models (VLMs), existing end-to-end VLA systems often lose key capabilities during fine-tuning as the model adapts to specific robotic tasks. We argue that a generalizable VLA model should retain and expand upon the VLM's core competencies: 1) Open-world embodied reasoning - the VLA should inherit the knowledge from VLM, i.e., recognize anything that the VLM can recognize, be capable of solving math problems, and possess visual-spatial intelligence, 2) Reasoning following - effectively translating the open-world reasoning into actionable steps for the robot. In this work, we introduce ChatVLA-2, a novel mixture-of-expert VLA model coupled with a specialized two-stage training pipeline designed to preserve the VLM's original strengths while enabling actionable reasoning. To validate our approach, we design a math-matching task wherein a robot interprets math problems written on a whiteboard and picks corresponding number cards from a table to solve equations. Remarkably, our method exhibits exceptional mathematical reasoning and OCR capabilities, despite these abilities not being explicitly trained within the VLA. Furthermore, we demonstrate that the VLA possesses strong spatial reasoning skills, enabling it to interpret novel directional instructions involving previously unseen objects. Overall, our method showcases reasoning and comprehension abilities that significantly surpass state-of-the-art imitation learning methods such as OpenVLA, DexVLA, and pi-zero. This work represents a substantial advancement toward developing truly generalizable robotic foundation models endowed with robust reasoning capacities.
Recent advances in diffusion$/$flow-matching policies have enabled imitation learning of complex, multi-modal action trajectories. However, they are computationally expensive because they sample a trajectory of trajectories: a diffusion$/$flow trajectory of action trajectories. They discard intermediate action trajectories, and must wait for the sampling process to complete before any actions can be executed on the robot. We simplify diffusion$/$flow policies by treating action trajectories as flow trajectories. Instead of starting from pure noise, our algorithm samples from a narrow Gaussian around the last action. Then, it incrementally integrates a velocity field learned via flow matching to produce a sequence of actions that constitute a single trajectory. This enables actions to be streamed to the robot on-the-fly during the flow sampling process, and is well-suited for receding horizon policy execution. Despite streaming, our method retains the ability to model multi-modal behavior. We train flows that stabilize around demonstration trajectories to reduce distribution shift and improve imitation learning performance. Streaming flow policy outperforms prior methods while enabling faster policy execution and tighter sensorimotor loops for learning-based robot control. Project website: https://streaming-flow-policy.github.io/
We study the problem of offline imitation learning in Markov decision processes (MDPs), where the goal is to learn a well-performing policy given a dataset of state-action pairs generated by an expert policy. Complementing a recent line of work on this topic that assumes the expert belongs to a tractable class of known policies, we approach this problem from a new angle and leverage a different type of structural assumption about the environment. Specifically, for the class of linear $Q^\pi$-realizable MDPs, we introduce a new algorithm called saddle-point offline imitation learning (\SPOIL), which is guaranteed to match the performance of any expert up to an additive error $\varepsilon$ with access to $\mathcal{O}(\varepsilon^{-2})$ samples. Moreover, we extend this result to possibly non-linear $Q^\pi$-realizable MDPs at the cost of a worse sample complexity of order $\mathcal{O}(\varepsilon^{-4})$. Finally, our analysis suggests a new loss function for training critic networks from expert data in deep imitation learning. Empirical evaluations on standard benchmarks demonstrate that the neural net implementation of \SPOIL is superior to behavior cloning and competitive with state-of-the-art algorithms.
Reinforcement learning (RL) is increasingly applied to real-world problems involving complex and structured decisions, such as routing, scheduling, and assortment planning. These settings challenge standard RL algorithms, which struggle to scale, generalize, and exploit structure in the presence of combinatorial action spaces. We propose Structured Reinforcement Learning (SRL), a novel actor-critic framework that embeds combinatorial optimization layers into the actor neural network. We enable end-to-end learning of the actor via Fenchel-Young losses and provide a geometric interpretation of SRL as a primal-dual algorithm in the dual of the moment polytope. Across six environments with exogenous and endogenous uncertainty, SRL matches or surpasses the performance of unstructured RL and imitation learning on static tasks and improves over these baselines by up to 92% on dynamic problems, with improved stability and convergence speed.
This work studies knowledge distillation (KD) for large language models (LLMs) through preference optimization. We propose a reward-guided imitation learning framework for sequential KD, formulating a min-max optimization problem between the policy and reward model (RM) to minimize the performance gap between the student and teacher policies. Specifically, the reward optimization is constrained to achieve near-optimality within a confidence set for preference alignment. For preference data construction, we explore both offline and online preference-based KD. Additionally, we reformulate the RM using the $Q$-value function and extend the framework to white-box KD, where the teacher policy's predicted probabilities are accessible. Theoretical analysis and empirical results demonstrate the effectiveness of the proposed framework.
Due to the powerful vision-language reasoning and generalization abilities, multimodal large language models (MLLMs) have garnered significant attention in the field of end-to-end (E2E) autonomous driving. However, their application to closed-loop systems remains underexplored, and current MLLM-based methods have not shown clear superiority to mainstream E2E imitation learning approaches. In this work, we propose ReasonPlan, a novel MLLM fine-tuning framework designed for closed-loop driving through holistic reasoning with a self-supervised Next Scene Prediction task and supervised Decision Chain-of-Thought process. This dual mechanism encourages the model to align visual representations with actionable driving context, while promoting interpretable and causally grounded decision making. We curate a planning-oriented decision reasoning dataset, namely PDR, comprising 210k diverse and high-quality samples. Our method outperforms the mainstream E2E imitation learning method by a large margin of 19% L2 and 16.1 driving score on Bench2Drive benchmark. Furthermore, ReasonPlan demonstrates strong zero-shot generalization on unseen DOS benchmark, highlighting its adaptability in handling zero-shot corner cases. Code and dataset will be found in https://github.com/Liuxueyi/ReasonPlan.
Controlling the flight of flapping-wing drones requires versatile controllers that handle their time-varying, nonlinear, and underactuated dynamics from incomplete and noisy sensor data. Model-based methods struggle with accurate modeling, while model-free approaches falter in efficiently navigating very high-dimensional and nonlinear control objective landscapes. This article presents a novel hybrid model-free/model-based approach to flight control based on the recently proposed reinforcement twinning algorithm. The model-based (MB) approach relies on an adjoint formulation using an adaptive digital twin, continuously identified from live trajectories, while the model-free (MF) approach relies on reinforcement learning. The two agents collaborate through transfer learning, imitation learning, and experience sharing using the real environment, the digital twin and a referee. The latter selects the best agent to interact with the real environment based on performance within the digital twin and a real-to-virtual environment consistency ratio. The algorithm is evaluated for controlling the longitudinal dynamics of a flapping-wing drone, with the environment simulated as a nonlinear, time-varying dynamical system under the influence of quasi-steady aerodynamic forces. The hybrid control learning approach is tested with three types of initialization of the adaptive model: (1) offline identification using previously available data, (2) random initialization with full online identification, and (3) offline pre-training with an estimation bias, followed by online adaptation. In all three scenarios, the proposed hybrid learning approach demonstrates superior performance compared to purely model-free and model-based methods.