Wenzhao Lian

Zheng Wu, Yichen Xie, Wenzhao Lian et al.

Humans are capable of abstracting various tasks as different combinations of multiple attributes. This perspective of compositionality is vital for human rapid learning and adaption since previous experiences from related tasks can be combined to generalize across novel compositional settings. In this work, we aim to achieve zero-shot policy generalization of Reinforcement Learning (RL) agents by leveraging the task compositionality. Our proposed method is a meta- RL algorithm with disentangled task representation, explicitly encoding different aspects of the tasks. Policy generalization is then performed by inferring unseen compositional task representations via the obtained disentanglement without extra exploration. The evaluation is conducted on three simulated tasks and a challenging real-world robotic insertion task. Experimental results demonstrate that our proposed method achieves policy generalization to unseen compositional tasks in a zero-shot manner.

Zixing Lei, Genjia Liu, Yuanshuo Zhang et al.

The field of Embodied AI is witnessing a rapid evolution toward general-purpose robotic systems, fueled by high-fidelity simulation and large-scale data collection. However, this scaling capability remains severely bottlenecked by a reliance on labor-intensive manual oversight from intricate reward shaping to hyperparameter tuning across heterogeneous backends. Inspired by LLMs' success in software automation and science discovery, we introduce \textsc{EmboCoach-Bench}, a benchmark evaluating the capacity of LLM agents to autonomously engineer embodied policies. Spanning 32 expert-curated RL and IL tasks, our framework posits executable code as the universal interface. We move beyond static generation to assess a dynamic closed-loop workflow, where agents leverage environment feedback to iteratively draft, debug, and optimize solutions, spanning improvements from physics-informed reward design to policy architectures such as diffusion policies. Extensive evaluations yield three critical insights: (1) autonomous agents can qualitatively surpass human-engineered baselines by 26.5\% in average success rate; (2) agentic workflow with environment feedback effectively strengthens policy development and substantially narrows the performance gap between open-source and proprietary models; and (3) agents exhibit self-correction capabilities for pathological engineering cases, successfully resurrecting task performance from near-total failures through iterative simulation-in-the-loop debugging. Ultimately, this work establishes a foundation for self-evolving embodied intelligence, accelerating the paradigm shift from labor-intensive manual tuning to scalable, autonomous engineering in embodied AI field.

Zhongxi Chen, Yifan Han, Yanming Shao et al.

Vision-Language-Action (VLA) models have emerged as a promising paradigm for grounding visual-language understanding into real-world robotic manipulation. However, dexterous manipulation remains challenging for VLA policies due to high-dimensional hand control and compounding execution errors, which makes real-world RL post-training essential for bridging the gap between visually grounded action generation and physically reliable dexterous execution. However, high-dimensional dexterous exploration often triggers temporal inconsistency, sample inefficiency and hardware risks in the real world. To address these challenges, we propose BORA, an offline-to-online RL post-training framework designed for real-world dexterous VLA models. In the offline phase, BORA constructs a critic that takes both the VLM's cognition tokens and action chunks as inputs. This design enables action-conditioned value guidance, allowing the critic to evaluate dexterous hand motions beyond visual context alone. During the subsequent online phase, BORA freezes the VLA base and introduces a lightweight, Human-in-the-Loop (HiL) chunk-wise residual adaptation mechanism to mitigate real-world execution errors and further correct the offline-learned intents within the actual physical environment. By inheriting the offline critic and employing intervention-driven rewards, BORA effectively corrects execution discrepancies and adapts to real-world physical variances while preserving the pretrained policy as a stable prior. Extensive evaluations across five complex real-world dexterous tasks demonstrate that BORA significantly outperforms pure imitation learning and traditional decoupled RL baselines, achieving a 33% absolute increase in average success rate under standard settings and up to a 43% improvement in unseen object generalization.

Yifan Han, Zhongxi Chen, Yuxuan Zhao et al.

While Vision-Language-Action (VLA) models have demonstrated promising generalization capabilities in robotic manipulation, deploying them on specific and complex downstream tasks still demands effective post-training. In parallel, Human-in-the-Loop (HiL) learning has proven to be a powerful mechanism for refining robot policies. However, extending this paradigm to dexterous manipulation remains challenging: multi-finger control is high-dimensional, contact-intensive, and exhibits execution distributions that differ markedly from standard arm motions, leaving existing dexterous VLA systems limited in reliability and adaptability. We present DexHiL, the first integrated arm-hand human-in-the-loop framework for dexterous VLA models, enabling coordinated interventions over the arm and the dexterous hand within a single system. DexHiL introduces an intervention-aware data sampling strategy that prioritizes corrective segments for post-training, alongside a lightweight teleoperation interface that supports instantaneous human corrections during execution. Real-robot experiments demonstrate that DexHiL serves as an effective post-training framework, yielding a substantial performance leap, outperforming standard offline-only fine-tuning baselines by an average of 25% in success rates across distinct tasks. Project page: https://chenzhongxi-sjtu.github.io/dexhil/

Yifan Han, Yichuan Peng, Pengfei Yi et al.

Dexterous grasp synthesis must jointly satisfy functional intent and physical feasibility, yet existing pipelines often decouple semantic grounding from refinement, yielding unstable or non-functional contacts under object and pose variations. This challenge is exacerbated by the high dimensionality and kinematic diversity of multi-fingered hands, which makes many methods rely on large, hardware-specific grasp datasets collected in simulation or through costly real-world trials. We propose a data-efficient framework that bypasses robot grasp data collection by exploiting object-centric semantic priors in pretrained generative diffusion models. Temporally aligned and fine-grained grasp affordances are extracted from raw human video demonstrations and fused with 3D scene geometry from depth images to infer semantically grounded contact targets. We further incorporate these affordance regions into the grasp refinement objective, explicitly guiding each fingertip toward its predicted region during optimization. The resulting system produces stable, human-intuitive multi-contact grasps across common objects and tools, while exhibiting strong generalization to previously unseen object instances within a category, pose variations, and multiple hand embodiments.This work (i) introduces a semantic affordance extraction pipeline leveraging vision--language generative priors for dexterous grasping, (ii) demonstrates cross-hand generalization without constructing hardware-specific grasp datasets, and (iii) establishes that a single depth modality suffices for high-performance grasp synthesis when coupled with foundation-model semantics. Our results highlight a path toward scalable, hardware-agnostic dexterous manipulation driven by human demonstrations and pretrained generative models.

Yifan Han, Jianxiang Liu, Haoyu Zhang et al.

Learning robot manipulation from human videos is appealing due to the scale and diversity of human demonstrations, but transferring such demonstrations to executable robot behavior remains challenging. Prior work either relies on robot data for downstream adaptation or learns affordance representations that remain at the perception level and do not directly support real-world execution. We present BridgeACT, an affordance-driven framework that learns robotic manipulation directly from human videos without requiring any robot demonstration data. Our key idea is to model affordance as an embodiment-agnostic intermediate representation that bridges human demonstrations and robot actions. BridgeACT decomposes manipulation into two complementary problems: where to grasp and how to move. To this end, BridgeACT first grounds task-relevant affordance regions in the current scene, and then predicts task-conditioned 3D motion affordances from human demonstrations. The resulting affordances are mapped to robot actions through a grasping module and a lightweight closed-loop motion controller, enabling direct deployment on real robots. In addition, we represent complex manipulation tasks as compositions of affordance operations, which allows a unified treatment of diverse tasks and object-to-object interactions. Experiments on real-world manipulation tasks show that BridgeACT outperforms prior baselines and generalizes to unseen objects, scenes, and viewpoints.

Jiashu Yang, Yifan Han, Yucheng Xie et al.

In embodied AI perception systems, visual perception should be active: the goal is not to passively process static images, but to actively acquire more informative data within pixel and spatial budget constraints. Existing vision models and fixed RGB-D camera systems fundamentally fail to reconcile wide-area coverage with fine-grained detail acquisition, severely limiting their efficacy in open-world robotic applications. To address this issue, we propose EyeVLA, a robotic eyeball for active visual perception that can take proactive actions based on instructions, enabling clear observation of fine-grained target objects and detailed information across a wide spatial extent. EyeVLA discretizes action behaviors into action tokens and integrates them with vision-language models (VLMs) that possess strong open-world understanding capabilities, enabling joint modeling of vision, language, and actions within a single autoregressive sequence. By using the 2D bounding box coordinates to guide the reasoning chain and applying reinforcement learning to refine the viewpoint selection policy, we transfer the open-world scene understanding capability of the VLM to a vision language action (VLA) policy using only minimal real-world data. Experiments show that our system efficiently performs instructed scenes in real-world environments and actively acquires more accurate visual information through instruction-driven actions of rotation and zoom, thereby achieving strong environmental perception capabilities. EyeVLA introduces a novel robotic vision system that leverages detailed and spatially rich, large-scale embodied data, and actively acquires highly informative visual observations for downstream embodied tasks.

Jialiang Li, Wenzheng Wu, Gaojing Zhang et al.

Successfully solving long-horizon manipulation tasks remains a fundamental challenge. These tasks involve extended action sequences and complex object interactions, presenting a critical gap between high-level symbolic planning and low-level continuous control. To bridge this gap, two essential capabilities are required: robust long-horizon task planning and effective goal-conditioned manipulation. Existing task planning methods, including traditional and LLM-based approaches, often exhibit limited generalization or sparse semantic reasoning. Meanwhile, image-conditioned control methods struggle to adapt to unseen tasks. To tackle these problems, we propose SAGE, a novel framework for Scene Graph-Aware Guidance and Execution in Long-Horizon Manipulation Tasks. SAGE utilizes semantic scene graphs as a structural representation for scene states. A structural scene graph enables bridging task-level semantic reasoning and pixel-level visuo-motor control. This also facilitates the controllable synthesis of accurate, novel sub-goal images. SAGE consists of two key components: (1) a scene graph-based task planner that uses VLMs and LLMs to parse the environment and reason about physically-grounded scene state transition sequences, and (2) a decoupled structural image editing pipeline that controllably converts each target sub-goal graph into a corresponding image through image inpainting and composition. Extensive experiments have demonstrated that SAGE achieves state-of-the-art performance on distinct long-horizon tasks.

Bowen Wen, Wenzhao Lian, Kostas Bekris et al.

Promising results have been achieved recently in category-level manipulation that generalizes across object instances. Nevertheless, it often requires expensive real-world data collection and manual specification of semantic keypoints for each object category and task. Additionally, coarse keypoint predictions and ignoring intermediate action sequences hinder adoption in complex manipulation tasks beyond pick-and-place. This work proposes a novel, category-level manipulation framework that leverages an object-centric, category-level representation and model-free 6 DoF motion tracking. The canonical object representation is learned solely in simulation and then used to parse a category-level, task trajectory from a single demonstration video. The demonstration is reprojected to a target trajectory tailored to a novel object via the canonical representation. During execution, the manipulation horizon is decomposed into longrange, collision-free motion and last-inch manipulation. For the latter part, a category-level behavior cloning (CatBC) method leverages motion tracking to perform closed-loop control. CatBC follows the target trajectory, projected from the demonstration and anchored to a dynamically selected category-level coordinate frame. The frame is automatically selected along the manipulation horizon by a local attention mechanism. This framework allows to teach different manipulation strategies by solely providing a single demonstration, without complicated manual programming. Extensive experiments demonstrate its efficacy in a range of challenging industrial tasks in highprecision assembly, which involve learning complex, long-horizon policies. The process exhibits robustness against uncertainty due to dynamics as well as generalization across object instances and scene configurations. The supplementary video is available at https://www.youtube.com/watch?v=WAr8ZY3mYyw

Bowen Wen, Wenzhao Lian, Kostas Bekris et al.

Task-relevant grasping is critical for industrial assembly, where downstream manipulation tasks constrain the set of valid grasps. Learning how to perform this task, however, is challenging, since task-relevant grasp labels are hard to define and annotate. There is also yet no consensus on proper representations for modeling or off-the-shelf tools for performing task-relevant grasps. This work proposes a framework to learn task-relevant grasping for industrial objects without the need of time-consuming real-world data collection or manual annotation. To achieve this, the entire framework is trained solely in simulation, including supervised training with synthetic label generation and self-supervised, hand-object interaction. In the context of this framework, this paper proposes a novel, object-centric canonical representation at the category level, which allows establishing dense correspondence across object instances and transferring task-relevant grasps to novel instances. Extensive experiments on task-relevant grasping of densely-cluttered industrial objects are conducted in both simulation and real-world setups, demonstrating the effectiveness of the proposed framework. Code and data are available at https://sites.google.com/view/catgrasp.

Jianlan Luo, Oleg Sushkov, Rugile Pevceviciute et al.

Over the past several years there has been a considerable research investment into learning-based approaches to industrial assembly, but despite significant progress these techniques have yet to be adopted by industry. We argue that it is the prohibitively large design space for Deep Reinforcement Learning (DRL), rather than algorithmic limitations per se, that are truly responsible for this lack of adoption. Pushing these techniques into the industrial mainstream requires an industry-oriented paradigm which differs significantly from the academic mindset. In this paper we define criteria for industry-oriented DRL, and perform a thorough comparison according to these criteria of one family of learning approaches, DRL from demonstration, against a professional industrial integrator on the recently established NIST assembly benchmark. We explain the design choices, representing several years of investigation, which enabled our DRL system to consistently outperform the integrator baseline in terms of both speed and reliability. Finally, we conclude with a competition between our DRL system and a human on a challenge task of insertion into a randomly moving target. This study suggests that DRL is capable of outperforming not only established engineered approaches, but the human motor system as well, and that there remains significant room for improvement. Videos can be found on our project website: https://sites.google.com/view/shield-nist.

Wenzhao Lian, Tim Kelch, Dirk Holz et al.

In recent years, many learning based approaches have been studied to realize robotic manipulation and assembly tasks, often including vision and force/tactile feedback. However, it remains frequently unclear what is the baseline state-of-the-art performance and what are the bottleneck problems. In this work, we evaluate some off-the-shelf (OTS) industrial solutions on a recently introduced benchmark, the National Institute of Standards and Technology (NIST) Assembly Task Boards. A set of assembly tasks are introduced and baseline methods are provided to understand their intrinsic difficulty. Multiple sensor-based robotic solutions are then evaluated, including hybrid force/motion control and 2D/3D pattern matching algorithms. An end-to-end integrated solution that accomplishes the tasks is also provided. The results and findings throughout the study reveal a few noticeable factors that impede the adoptions of the OTS solutions: expertise dependent, limited applicability, lack of interoperability, no scene awareness or error recovery mechanisms, and high cost. This paper also provides a first attempt of an objective benchmark performance on the NIST Assembly Task Boards as a reference comparison for future works on this problem.

Peter A. Zachares, Michelle A. Lee, Wenzhao Lian et al.

A key challenge towards the goal of multi-part assembly tasks is finding robust sensorimotor control methods in the presence of uncertainty. In contrast to previous works that rely on a priori knowledge on whether two parts match, we aim to learn this through physical interaction. We propose a hierarchical approach that enables a robot to autonomously assemble parts while being uncertain about part types and positions. In particular, our probabilistic approach learns a set of differentiable filters that leverage the tactile sensorimotor trace from failed assembly attempts to update its belief about part position and type. This enables a robot to overcome assembly failure. We demonstrate the effectiveness of our approach on a set of object fitting tasks. The experimental results indicate that our proposed approach achieves higher precision in object position and type estimation, and accomplishes object fitting tasks faster than baselines.

Zheng Wu, Wenzhao Lian, Vaibhav Unhelkar et al.

Rewards play a crucial role in reinforcement learning. To arrive at the desired policy, the design of a suitable reward function often requires significant domain expertise as well as trial-and-error. Here, we aim to minimize the effort involved in designing reward functions for contact-rich manipulation tasks. In particular, we provide an approach capable of extracting dense reward functions algorithmically from robots' high-dimensional observations, such as images and tactile feedback. In contrast to state-of-the-art high-dimensional reward learning methodologies, our approach does not leverage adversarial training, and is thus less prone to the associated training instabilities. Instead, our approach learns rewards by estimating task progress in a self-supervised manner. We demonstrate the effectiveness and efficiency of our approach on two contact-rich manipulation tasks, namely, peg-in-hole and USB insertion. The experimental results indicate that the policies trained with the learned reward function achieves better performance and faster convergence compared to the baselines.

Makoto Yamada, Wenzhao Lian, Amit Goyal et al.

We propose the convex factorization machine (CFM), which is a convex variant of the widely used Factorization Machines (FMs). Specifically, we employ a linear+quadratic model and regularize the linear term with the $\ell_2$-regularizer and the quadratic term with the trace norm regularizer. Then, we formulate the CFM optimization as a semidefinite programming problem and propose an efficient optimization procedure with Hazan's algorithm. A key advantage of CFM over existing FMs is that it can find a globally optimal solution, while FMs may get a poor locally optimal solution since the objective function of FMs is non-convex. In addition, the proposed algorithm is simple yet effective and can be implemented easily. Finally, CFM is a general factorization method and can also be used for other factorization problems including including multi-view matrix factorization and tensor completion problems. Through synthetic and movielens datasets, we first show that the proposed CFM achieves results competitive to FMs. Furthermore, in a toxicogenomics prediction task, we show that CFM outperforms a state-of-the-art tensor factorization method.