LatentSightDrive: Progressive Foresight Internalization for Autonomous Driving

Xue Zhao, Yiming Zhong, Zemin Yang, Xiang Feng, Jin Pan, Xinbing Wang, Xinge Zhu, Yuexin Ma†, Nanyang Ye†

We introduce LatentSightDrive, a framework that internalizes future evidence from an external world model for autonomous driving. Scene-adaptive guidance and planning-relevant foresight scoring selectively align internal and external latent representations to support trajectory planning.

Implicit Drifting Policy: One-Step Action Generation via Conditional Expert Geometry

Zemin Yang, Yaoyu He, Yiming Zhong, Yuhao Zhang, Xinge Zhu, Yao Mu, Qingqiu Huang, Yuexin Ma†

We introduce Implicit Drifting Policy, a one-step imitation learning framework that uses conditional expert geometry to guide policy training without explicit vector field estimation. It combines efficient action generation with geometric constraints, achieving competitive performance across 2D, 3D, and real-world manipulation tasks.

ResVLA: From Noise to Intent: Anchoring Generative VLA Policies with Residual Bridges

Yiming Zhong*, Yaoyu He*, Zemin Yang*, Pengfei Tian, Yifan Huang, Qingqiu Huang, Xinge Zhu, Yuexin Ma†

We introduce ResVLA, a generative vision-language-action framework that shifts robot control from generation-from-noise to refinement-from-intent by anchoring low-frequency semantic intent and refining high-frequency residual dynamics, achieving strong robustness, faster convergence, and competitive performance.

FreqPolicy: Frequency Autoregressive Visuomotor Policy with Continuous Tokens

Yiming Zhong, Yumeng Liu, Chuyang Xiao, Zemin Yang, Youzhuo Wang, Yufei Zhu, Ye Shi, Yujing Sun, Xinge Zhu, Yuexin Ma†

This paper proposes FreqPolicy, a frequency-domain autoregressive visuomotor policy that progressively models hierarchical frequency components with continuous latent representations, achieving superior accuracy and efficiency in robotic manipulation tasks.

FastGrasp: Learning-based Whole-body Control method for Fast Dexterous Grasping with Mobile Manipulators

Heng Tao*, Yiming Zhong*, Zemin Yang*, Yuexin Ma†

We introduce FastGrasp, a learning-based framework that combines grasp guidance, whole-body control, and tactile feedback for fast dexterous mobile manipulation. A two-stage reinforcement learning pipeline coordinates the mobile base, arm, and hand, enabling robust grasping in simulation and the real world.

DexH2R: A Benchmark for Dynamic Dexterous Grasping in Human-to-Robot Handover

Youzhuo Wang*, Jiayi Ye*, Chuyang Xiao, Yiming Zhong, Heng Tao, Hang Yu, Yumeng Liu, Jingyi Yu, Yuexin Ma†

This paper introduces DexH2R, a real-world dataset for human-to-robot handovers featuring dexterous motions, diverse objects, and rich annotations. Using teleoperation, it captures natural human-like behaviors for robotic learning.

EvolvingGrasp: Evolutionary Grasp Generation via Efficient Preference Alignment

Yufei Zhu*, Yiming Zhong*, Zemin Yang, Peishan Cong, Jingyi Yu, Xinge Zhu, Yuexin Ma†

This paper introduces EvolvingGrasp, which integrates Handpose-wise Preference Optimization with a Physics-aware Consistency Model to enable efficient evolutionary grasp generation, achieving improved grasp success rates and computational efficiency.

DexGraspAnything: Towards Universal Robotic Dexterous Grasping with Physics Awareness

Yiming Zhong*, Qi Jiang*, Jingyi Yu, Yuexin Ma†

This paper proposes DexGrasp Anything, a diffusion-based method for generating physically plausible grasps with dexterous hands. By integrating physical constraints into both training and sampling, we address high-DOF challenges while synthesizing robust poses for diverse objects. Our 3.4M-grasp dataset (15k+ objects) enables scalable learning, achieving state-of-the-art performance in universal robotic grasping across benchmarks.

UniAfford: Token-Routed Multitask Learning for Generalizable 2D-3D Affordance Perception

Yuhao Liu, Yiming Zhong‡, Hanqing Wang, Shaocheng Yan, Yuhang Zhang, Wenzhou Lyu, Ziyang Ding, Wei Zhang, Xue Chao, Jin Pan, Yuexin Ma†, Xinge Zhu†

We introduce UniAfford, a unified framework for generalizable 2D-3D affordance perception. A shared multimodal language model uses token-based task routing and modality-specific decoders to learn from pixel-level and point-level supervision, supporting image, point-cloud, and joint multimodal inputs.

VideoAfford: Grounding 3D Affordance from Human-Object-Interaction Videos via Multimodal Large Language Model

Hanqing Wang, Mingyu Liu, Xiaoyu Chen, Chengwei Ma, Yiming Zhong, Wenti Yin, Yuhao Liu, Zhiqing Cui, Jiahao Yuan, Lu Dai, Zhiyuan Ma†, Hui Xiong†

We introduce VideoAfford and the VIDA dataset to learn 3D affordances from human-object interaction videos. By combining multimodal language models with latent action priors and a spatial-aware loss, VideoAfford enables fine-grained affordance grounding and reasoning with strong open-world generalization.

Affordance-R1: Reinforcement Learning for Generalizable Affordance Reasoning in Multimodal Large Language Model

Hanqing Wang*, Shaoyang Wang*, Yiming Zhong, Zemin Yang, Jiamin Wang, Zhiqing Cui,Jiahao Yuan, Yifan Han, Mingyu Liu, Yuexin Ma†

We introduce Affordance-R1, which is capable of generating explicit reasoning alongside the final answer. With the help of proposed affordance reasoning reward, it achieves robust zero-shot generalization and exhibits emergent test-time reasoning capabilities.