Jingkai Sun

Jiahang Cao, Qiang Zhang, Jingkai Sun et al.

Diffusion models have been widely employed in the field of 3D manipulation due to their efficient capability to learn distributions, allowing for precise prediction of action trajectories. However, diffusion models typically rely on large parameter UNet backbones as policy networks, which can be challenging to deploy on resource-constrained devices. Recently, the Mamba model has emerged as a promising solution for efficient modeling, offering low computational complexity and strong performance in sequence modeling. In this work, we propose the Mamba Policy, a lighter but stronger policy that reduces the parameter count by over 80% compared to the original policy network while achieving superior performance. Specifically, we introduce the XMamba Block, which effectively integrates input information with conditional features and leverages a combination of Mamba and Attention mechanisms for deep feature extraction. Extensive experiments demonstrate that the Mamba Policy excels on the Adroit, Dexart, and MetaWorld datasets, requiring significantly fewer computational resources. Additionally, we highlight the Mamba Policy's enhanced robustness in long-horizon scenarios compared to baseline methods and explore the performance of various Mamba variants within the Mamba Policy framework. Real-world experiments are also conducted to further validate its effectiveness. Our open-source project page can be found at https://andycao1125.github.io/mamba_policy/.

Jiaxu Wang, Junhao He, Jingkai Sun et al.

Learning real-world dynamics from visual observations is crucial for various domains. A common strategy is to calibrate simulators by estimating physical parameters, yet accuracy is ultimately bounded by the underlying physical models, which often assume materials are homogeneous and isotropic. Even if reasonable, real-world objects typically exhibit mild anisotropy and heterogeneity. After the near-isotropic backbone is well calibrated, these residual effects become the key bottleneck for further closing the real-to-sim gap. Although neural networks can fit dynamics end-to-end, such black-box modeling discards strong physical priors, leading to poor data efficiency and overfitting. Therefore, we propose MoSA, a motion-constrained stress adaptation framework that targets these residual effects to further improve real-to-sim dynamics learning. MoSA uses an isotropic model as a physics prior and learns residual stress operators to capture mild anisotropy and heterogeneity. It progressively adapts stresses via microplane-constrained redistribution in a physics-informed cascaded network. We further impose motion constraints by supervising temporal and spatial derivatives of the deformation field. Experimentally, our learned dynamics achieves superior accuracy, generalization, and robustness, while learning physically meaningful residual anisotropy. Finally, we validate MoSA in a robot manipulation setting, showing that better real-to-sim dynamics modeling translates into more reliable sim-to-real transfer. Project Page is available at https://mercerai.github.io/MoSA/.

Xiaoyang Jiang, Qiang Zhang, Jingkai Sun et al.

Recent advancements in legged robots using deep reinforcement learning have led to significant progress. Quadruped robots can perform complex tasks in challenging environments, while bipedal and humanoid robots have also achieved breakthroughs. Current reinforcement learning methods leverage diverse robot bodies and historical information to perform actions, but previous research has not emphasized the speed and energy consumption of network inference and the biological significance of neural networks. Most networks are traditional artificial neural networks that utilize multilayer perceptrons (MLP). This paper presents a novel Spiking Neural Network (SNN) for legged robots, showing exceptional performance in various simulated terrains. SNNs provide natural advantages in inference speed and energy consumption, and their pulse-form processing enhances biological interpretability. This study presents a highly efficient SNN for legged robots that can be seamless integrated into other learning models.

Zelin Tao, Zeran Su, Peiran Liu et al.

Achieving general-purpose humanoid control requires a delicate balance between the precise execution of commanded motions and the flexible, anthropomorphic adaptability needed to recover from unpredictable environmental perturbations. Current general controllers predominantly formulate motion control as a rigid reference-tracking problem. While effective in nominal conditions, these trackers often exhibit brittle, non-anthropomorphic failure modes under severe disturbances, lacking the generative adaptability inherent to human motor control. To overcome this limitation, we propose Heracles, a novel state-conditioned diffusion middleware that bridges precise motion tracking and generative synthesis. Rather than relying on rigid tracking paradigms or complex explicit mode-switching, Heracles operates as an intermediary layer between high-level reference motions and low-level physics trackers. By conditioning on the robot's real-time state, the diffusion model implicitly adapts its behavior: it approximates an identity map when the state closely aligns with the reference, preserving zero-shot tracking fidelity. Conversely, when encountering significant state deviations, it seamlessly transitions into a generative synthesizer to produce natural, anthropomorphic recovery trajectories. Our framework demonstrates that integrating generative priors into the control loop not only significantly enhances robustness against extreme perturbations but also elevates humanoid control from a rigid tracking paradigm to an open-ended, generative general-purpose architecture.

Yicheng Jiang, Jiaxu Wang, Junhao He et al.

Current 3D-aware pretraining methods for embodied perception and manipulation are largely built on differentiable rendering frameworks, producing either fully implicit neural fields or fully explicit geometric primitives. Implicit representations, while expressive, lack explicit structural cues, whereas explicit ones preserve geometry but suffer from resolution limits and weak generalization. To address these limitations, we propose a novel pretraining framework that learns a hybrid representation-structural latent points. Specifically, we insert a point-wise latent variational autoencoder into the latent space of a point-cloud autoencoder, jointly regularizing point-wise features and coordinates toward a Gaussian prior. The resulting compact latent preserves coarse structural tendencies, which do not encode precise geometry but capture richer rough shape and semantic information, effectively combining the expressiveness of implicit representations with the structural priors of explicit ones. In addition, informed by shared design choices in prior work, we develop a streamlined, efficient 3DGS-based rendering pipeline that is deliberately kept lightweight, improving efficiency while leaving greater representational capacity to the front-end latent module. Extensive evaluations on RLBench, ManiSkill2, and a real-robot platform demonstrate consistent gains in task success, sample efficiency, and robustness to viewpoint and scene variations over strong baselines. Ablation studies further confirm that each component of our framework is critical to overall performance.

Pengfei Yao, Tianlu Mao, Min Shi et al.

Trajectory prediction plays a crucial role in autonomous driving. Existing mainstream research and continuoual learning-based methods all require training on complete datasets, leading to poor prediction accuracy when sudden changes in scenarios occur and failing to promptly respond and update the model. Whether these methods can make a prediction in real-time and use data instances to update the model immediately(i.e., online learning settings) remains a question. The problem of gradient explosion or vanishing caused by data instance streams also needs to be addressed. Inspired by Hedge Propagation algorithm, we propose Expert Attention Network, a complete online learning framework for trajectory prediction. We introduce expert attention, which adjusts the weights of different depths of network layers, avoiding the model updated slowly due to gradient problem and enabling fast learning of new scenario's knowledge to restore prediction accuracy. Furthermore, we propose a short-term motion trend kernel function which is sensitive to scenario change, allowing the model to respond quickly. To the best of our knowledge, this work is the first attempt to address the online learning problem in trajectory prediction. The experimental results indicate that traditional methods suffer from gradient problems and that our method can quickly reduce prediction errors and reach the state-of-the-art prediction accuracy.

Jiaxu Wang, Yicheng Jiang, Tianlun He et al.

World-model-based imagine-then-act becomes a promising paradigm for robotic manipulation, yet existing approaches typically support either purely image-based forecasting or reasoning over partial 3D geometry, limiting their ability to predict complete 4D scene dynamics. This work proposes a novel embodied 4D world model that enables geometrically consistent, arbitrary-view RGBD generation: given only a single-view RGBD observation as input, the model imagines the remaining viewpoints, which can then be back-projected and fused to assemble a more complete 3D structure across time. To efficiently learn the multi-view, cross-modality generation, we explicitly design cross-view and cross-modality feature fusion that jointly encourage consistency between RGB and depth and enforce geometric alignment across views. Beyond prediction, converting generated futures into actions is often handled by inverse dynamics, which is ill-posed because multiple actions can explain the same transition. We address this with a test-time action optimization strategy that backpropagates through the generative model to infer a trajectory-level latent best matching the predicted future, and a residual inverse dynamics model that turns this trajectory prior into accurate executable actions. Experiments on three datasets demonstrate strong performance on both 4D scene generation and downstream manipulation, and ablations provide practical insights into the key design choices.

Qiang Zhang, Jiahao Ma, Peiran Liu et al.

Humanoid motion control has witnessed significant breakthroughs in recent years, with deep reinforcement learning (RL) emerging as a primary catalyst for achieving complex, human-like behaviors. However, the high dimensionality and intricate dynamics of humanoid robots make manual motion design impractical, leading to a heavy reliance on expensive motion capture (MoCap) data. These datasets are not only costly to acquire but also frequently lack the necessary geometric context of the surrounding physical environment. Consequently, existing motion synthesis frameworks often suffer from a decoupling of motion and scene, resulting in physical inconsistencies such as contact slippage or mesh penetration during terrain-aware tasks. In this work, we present MeshMimic, an innovative framework that bridges 3D scene reconstruction and embodied intelligence to enable humanoid robots to learn coupled "motion-terrain" interactions directly from video. By leveraging state-of-the-art 3D vision models, our framework precisely segments and reconstructs both human trajectories and the underlying 3D geometry of terrains and objects. We introduce an optimization algorithm based on kinematic consistency to extract high-quality motion data from noisy visual reconstructions, alongside a contact-invariant retargeting method that transfers human-environment interaction features to the humanoid agent. Experimental results demonstrate that MeshMimic achieves robust, highly dynamic performance across diverse and challenging terrains. Our approach proves that a low-cost pipeline utilizing only consumer-grade monocular sensors can facilitate the training of complex physical interactions, offering a scalable path toward the autonomous evolution of humanoid robots in unstructured environments.

Yeke Chen, Shihao Dong, Xiaoyu Ji et al.

Realizing versatile and human-like performance in high-demand sports like badminton remains a formidable challenge for humanoid robotics. Unlike standard locomotion or static manipulation, this task demands a seamless integration of explosive whole-body coordination and precise, timing-critical interception. While recent advances have achieved lifelike motion mimicry, bridging the gap between kinematic imitation and functional, physics-aware striking without compromising stylistic naturalness is non-trivial. To address this, we propose Imitation-to-Interaction, a progressive reinforcement learning framework designed to evolve a robot from a "mimic" to a capable "striker." Our approach establishes a robust motor prior from human data, distills it into a compact, model-based state representation, and stabilizes dynamics via adversarial priors. Crucially, to overcome the sparsity of expert demonstrations, we introduce a manifold expansion strategy that generalizes discrete strike points into a dense interaction volume. We validate our framework through the mastery of diverse skills, including lifts and drop shots, in simulation. Furthermore, we demonstrate the first zero-shot sim-to-real transfer of anthropomorphic badminton skills to a humanoid robot, successfully replicating the kinetic elegance and functional precision of human athletes in the physical world.

Jiaxu Wang, Junhao He, Ziyi Zhang et al.

Event cameras offer promising advantages such as high dynamic range and low latency, making them well-suited for challenging lighting conditions and fast-moving scenarios. However, reconstructing 3D scenes from raw event streams is difficult because event data is sparse and does not carry absolute color information. To release its potential in 3D reconstruction, we propose the first event-based generalizable 3D reconstruction framework, called EvGGS, which reconstructs scenes as 3D Gaussians from only event input in a feedforward manner and can generalize to unseen cases without any retraining. This framework includes a depth estimation module, an intensity reconstruction module, and a Gaussian regression module. These submodules connect in a cascading manner, and we collaboratively train them with a designed joint loss to make them mutually promote. To facilitate related studies, we build a novel event-based 3D dataset with various material objects and calibrated labels of grayscale images, depth maps, camera poses, and silhouettes. Experiments show models that have jointly trained significantly outperform those trained individually. Our approach performs better than all baselines in reconstruction quality, and depth/intensity predictions with satisfactory rendering speed.

Jiaxu Wang, Qiang Zhang, Jingkai Sun et al.

An excellent representation is crucial for reinforcement learning (RL) performance, especially in vision-based reinforcement learning tasks. The quality of the environment representation directly influences the achievement of the learning task. Previous vision-based RL typically uses explicit or implicit ways to represent environments, such as images, points, voxels, and neural radiance fields. However, these representations contain several drawbacks. They cannot either describe complex local geometries or generalize well to unseen scenes, or require precise foreground masks. Moreover, these implicit neural representations are akin to a ``black box", significantly hindering interpretability. 3D Gaussian Splatting (3DGS), with its explicit scene representation and differentiable rendering nature, is considered a revolutionary change for reconstruction and representation methods. In this paper, we propose a novel Generalizable Gaussian Splatting framework to be the representation of RL tasks, called GSRL. Through validation in the RoboMimic environment, our method achieves better results than other baselines in multiple tasks, improving the performance by 10%, 44%, and 15% compared with baselines on the hardest task. This work is the first attempt to leverage generalizable 3DGS as a representation for RL.

Yi Zhang, Qiang Zhang, Xiaozhu Ju et al.

While multimodal large language models (MLLMs) have made groundbreaking progress in embodied intelligence, they still face significant challenges in spatial reasoning for complex long-horizon tasks. To address this gap, we propose EmbodiedVSR (Embodied Visual Spatial Reasoning), a novel framework that integrates dynamic scene graph-guided Chain-of-Thought (CoT) reasoning to enhance spatial understanding for embodied agents. By explicitly constructing structured knowledge representations through dynamic scene graphs, our method enables zero-shot spatial reasoning without task-specific fine-tuning. This approach not only disentangles intricate spatial relationships but also aligns reasoning steps with actionable environmental dynamics. To rigorously evaluate performance, we introduce the eSpatial-Benchmark, a comprehensive dataset including real-world embodied scenarios with fine-grained spatial annotations and adaptive task difficulty levels. Experiments demonstrate that our framework significantly outperforms existing MLLM-based methods in accuracy and reasoning coherence, particularly in long-horizon tasks requiring iterative environment interaction. The results reveal the untapped potential of MLLMs for embodied intelligence when equipped with structured, explainable reasoning mechanisms, paving the way for more reliable deployment in real-world spatial applications. The codes and datasets will be released soon.

Zhang Zhang, Qiang Zhang, Wei Cui et al.

Understanding and forecasting the scene evolutions deeply affect the exploration and decision of embodied agents. While traditional methods simulate scene evolutions through trajectory prediction of potential instances, current works use the occupancy world model as a generative framework for describing fine-grained overall scene dynamics. However, existing methods cluster on the outdoor structured road scenes, while ignoring the exploration of forecasting 3D occupancy scene evolutions for robots in indoor scenes. In this work, we explore a new framework for learning the scene evolutions of observed fine-grained occupancy and propose an occupancy world model based on the combined spatio-temporal receptive field and guided autoregressive transformer to forecast the scene evolutions, called RoboOccWorld. We propose the Conditional Causal State Attention (CCSA), which utilizes camera poses of next state as conditions to guide the autoregressive transformer to adapt and understand the indoor robotics scenarios. In order to effectively exploit the spatio-temporal cues from historical observations, Hybrid Spatio-Temporal Aggregation (HSTA) is proposed to obtain the combined spatio-temporal receptive field based on multi-scale spatio-temporal windows. In addition, we restructure the OccWorld-ScanNet benchmark based on local annotations to facilitate the evaluation of the indoor 3D occupancy scene evolution prediction task. Experimental results demonstrate that our RoboOccWorld outperforms state-of-the-art methods in indoor 3D occupancy scene evolution prediction task. The code will be released soon.

Wei Cui, Haoyu Wang, Wenkang Qin et al.

Humanoid robot technology is advancing rapidly, with manufacturers introducing diverse heterogeneous visual perception modules tailored to specific scenarios. Among various perception paradigms, occupancy-based representation has become widely recognized as particularly suitable for humanoid robots, as it provides both rich semantic and 3D geometric information essential for comprehensive environmental understanding. In this work, we present Humanoid Occupancy, a generalized multimodal occupancy perception system that integrates hardware and software components, data acquisition devices, and a dedicated annotation pipeline. Our framework employs advanced multi-modal fusion techniques to generate grid-based occupancy outputs encoding both occupancy status and semantic labels, thereby enabling holistic environmental understanding for downstream tasks such as task planning and navigation. To address the unique challenges of humanoid robots, we overcome issues such as kinematic interference and occlusion, and establish an effective sensor layout strategy. Furthermore, we have developed the first panoramic occupancy dataset specifically for humanoid robots, offering a valuable benchmark and resource for future research and development in this domain. The network architecture incorporates multi-modal feature fusion and temporal information integration to ensure robust perception. Overall, Humanoid Occupancy delivers effective environmental perception for humanoid robots and establishes a technical foundation for standardizing universal visual modules, paving the way for the widespread deployment of humanoid robots in complex real-world scenarios.

Jingkai Sun, Qiang Zhang, Jiaxu Wang et al.

Dynamic vision sensors (DVS) are bio-inspired devices that capture visual information in the form of asynchronous events, which encode changes in pixel intensity with high temporal resolution and low latency. These events provide rich motion cues that can be exploited for various computer vision tasks, such as action recognition. However, most existing DVS-based action recognition methods lose temporal information during data transformation or suffer from noise and outliers caused by sensor imperfections or environmental factors. To address these challenges, we propose a novel framework that preserves and exploits the spatiotemporal structure of event data for action recognition. Our framework consists of two main components: 1) a point-wise event masked autoencoder (MAE) that learns a compact and discriminative representation of event patches by reconstructing them from masked raw event camera points data; 2) an improved event points patch generation algorithm that leverages an event data inlier model and point-wise data augmentation techniques to enhance the quality and diversity of event points patches. To the best of our knowledge, our approach introduces the pre-train method into event camera raw points data for the first time, and we propose a novel event points patch embedding to utilize transformer-based models on event cameras.

Jiaxu Wang, Jingkai Sun, Junhao He et al.

Learning-based simulators show great potential for simulating particle dynamics when 3D groundtruth is available, but per-particle correspondences are not always accessible. The development of neural rendering presents a new solution to this field to learn 3D dynamics from 2D images by inverse rendering. However, existing approaches still suffer from ill-posed natures resulting from the 2D to 3D uncertainty, for example, specific 2D images can correspond with various 3D particle distributions. To mitigate such uncertainty, we consider a conventional, mechanically interpretable framework as the physical priors and extend it to a learning-based version. In brief, we incorporate the learnable graph kernels into the classic Discrete Element Analysis (DEA) framework to implement a novel mechanics-integrated learning system. In this case, the graph network kernels are only used for approximating some specific mechanical operators in the DEA framework rather than the whole dynamics mapping. By integrating the strong physics priors, our methods can effectively learn the dynamics of various materials from the partial 2D observations in a unified manner. Experiments show that our approach outperforms other learned simulators by a large margin in this context and is robust to different renderers, fewer training samples, and fewer camera views.

Qiang Zhang, Jiahang Cao, Jingkai Sun et al.

In recent years, quadruped robotics has advanced significantly, particularly in perception and motion control via reinforcement learning, enabling complex motions in challenging environments. Visual sensors like depth cameras enhance stability and robustness but face limitations, such as low operating frequencies relative to joint control and sensitivity to lighting, which hinder outdoor deployment. Additionally, deep neural networks in sensor and control systems increase computational demands. To address these issues, we introduce spiking neural networks (SNNs) and event cameras to perform a challenging quadruped parkour task. Event cameras capture dynamic visual data, while SNNs efficiently process spike sequences, mimicking biological perception. Experimental results demonstrate that this approach significantly outperforms traditional models, achieving excellent parkour performance with just 11.7% of the energy consumption of an artificial neural network (ANN)-based model, yielding an 88.3% energy reduction. By integrating event cameras with SNNs, our work advances robotic reinforcement learning and opens new possibilities for applications in demanding environments.

Junhao He, Jiaxu Wang, Jia Li et al.

Reconstructing Dynamic 3D Gaussian Splatting (3DGS) from low-framerate RGB videos is challenging. This is because large inter-frame motions will increase the uncertainty of the solution space. For example, one pixel in the first frame might have more choices to reach the corresponding pixel in the second frame. Event cameras can asynchronously capture rapid visual changes and are robust to motion blur, but they do not provide color information. Intuitively, the event stream can provide deterministic constraints for the inter-frame large motion by the event trajectories. Hence, combining low-temporal-resolution images with high-framerate event streams can address this challenge. However, it is challenging to jointly optimize Dynamic 3DGS using both RGB and event modalities due to the significant discrepancy between these two data modalities. This paper introduces a novel framework that jointly optimizes dynamic 3DGS from the two modalities. The key idea is to adopt event motion priors to guide the optimization of the deformation fields. First, we extract the motion priors encoded in event streams by using the proposed LoCM unsupervised fine-tuning framework to adapt an event flow estimator to a certain unseen scene. Then, we present the geometry-aware data association method to build the event-Gaussian motion correspondence, which is the primary foundation of the pipeline, accompanied by two useful strategies, namely motion decomposition and inter-frame pseudo-label. Extensive experiments show that our method outperforms existing image and event-based approaches across synthetic and real scenes and prove that our method can effectively optimize dynamic 3DGS with the help of event data.

Jiahang Cao, Yize Huang, Hanzhong Guo et al.

Diffusion-based models for robotic control, including vision-language-action (VLA) and vision-action (VA) policies, have demonstrated significant capabilities. Yet their advancement is constrained by the high cost of acquiring large-scale interaction datasets. This work introduces an alternative paradigm for enhancing policy performance without additional model training. Perhaps surprisingly, we demonstrate that the composed policies can exceed the performance of either parent policy. Our contribution is threefold. First, we establish a theoretical foundation showing that the convex composition of distributional scores from multiple diffusion models can yield a superior one-step functional objective compared to any individual score. A Grönwall-type bound is then used to show that this single-step improvement propagates through entire generation trajectories, leading to systemic performance gains. Second, motivated by these results, we propose General Policy Composition (GPC), a training-free method that enhances performance by combining the distributional scores of multiple pre-trained policies via a convex combination and test-time search. GPC is versatile, allowing for the plug-and-play composition of heterogeneous policies, including VA and VLA models, as well as those based on diffusion or flow-matching, irrespective of their input visual modalities. Third, we provide extensive empirical validation. Experiments on Robomimic, PushT, and RoboTwin benchmarks, alongside real-world robotic evaluations, confirm that GPC consistently improves performance and adaptability across a diverse set of tasks. Further analysis of alternative composition operators and weighting strategies offers insights into the mechanisms underlying the success of GPC. These results establish GPC as a simple yet effective method for improving control performance by leveraging existing policies.

Haozhuo Zhang, Jingkai Sun, Michele Caprio et al.

We introduce HumanoidVerse, a novel framework for vision-language guided humanoid control that enables a single physically simulated robot to perform long-horizon, multi-object rearrangement tasks across diverse scenes. Unlike prior methods that operate in fixed settings with single-object interactions, our approach supports consecutive manipulation of multiple objects, guided only by natural language instructions and egocentric camera RGB observations. HumanoidVerse is trained via a multi-stage curriculum using a dual-teacher distillation pipeline, enabling fluid transitions between sub-tasks without requiring environment resets. To support this, we construct a large-scale dataset comprising 350 multi-object tasks spanning four room layouts. Extensive experiments in the Isaac Gym simulator demonstrate that our method significantly outperforms prior state-of-the-art in both task success rate and spatial precision, and generalizes well to unseen environments and instructions. Our work represents a key step toward robust, general-purpose humanoid agents capable of executing complex, sequential tasks under real-world sensory constraints. The video visualization results can be found on the project page: https://haozhuo-zhang.github.io/HumanoidVerse-project-page/.

Jiahang Cao, Qiang Zhang, Ziqing Wang et al.

Sequential modeling has demonstrated remarkable capabilities in offline reinforcement learning (RL), with Decision Transformer (DT) being one of the most notable representatives, achieving significant success. However, RL trajectories possess unique properties to be distinguished from the conventional sequence (e.g., text or audio): (1) local correlation, where the next states in RL are theoretically determined solely by current states and actions based on the Markov Decision Process (MDP), and (2) global correlation, where each step's features are related to long-term historical information due to the time-continuous nature of trajectories. In this paper, we propose a novel action sequence predictor, named Mamba Decision Maker (MambaDM), where Mamba is expected to be a promising alternative for sequence modeling paradigms, owing to its efficient modeling of multi-scale dependencies. In particular, we introduce a novel mixer module that proficiently extracts and integrates both global and local features of the input sequence, effectively capturing interrelationships in RL datasets. Extensive experiments demonstrate that MambaDM achieves state-of-the-art performance in Atari and OpenAI Gym datasets. Furthermore, we empirically investigate the scaling laws of MambaDM, finding that increasing model size does not bring performance improvement, but scaling the dataset amount by 2x for MambaDM can obtain up to 33.7% score improvement on Atari dataset. This paper delves into the sequence modeling capabilities of MambaDM in the RL domain, paving the way for future advancements in robust and efficient decision-making systems.