Motus2: A Self-Evolving General World Model for Dexterous Manipulation

Hongzhe Bi, Zihao Zhou, Yihang Tang, Jingrui Pang, Shuhe Huang, Haitian Liu, Runqing Wang, Shuai Huang, Yichen Wang, Yiming Cheng, Ruowen Zhao, Zhenghua Li, Hengkai Tan, Xiaolong Liu, Jinhui Wan, Jiabao Liu, Min Zhao, Fan Bao, Jun Zhu

Project

• Motus2 is the first fully multimodal General World Model for high-DoF dexterous manipulation, integrating vision, language, action, and touch. It is the first to realize closed-loop self-evolution with one shared model serving as a policy, simulator, and evaluator. Expert demonstrations supervise action learning, while failed and suboptimal interactions support dynamics and value learning. Data scaling spans approximately 130,000 raw hours of monocular and stereo egocentric recordings, followed by robot-domain adaptation at the hundred-hour scale. The model further explores long-context modeling and adds tactile feedback for contact-aware control. Together, egocentric data scaling and closed-loop general world model scaling provide a general path toward self-evolving dexterous manipulation.

Motus: A Unified Latent Action World Model

Hongzhe Bi, Hengkai Tan, Shenghao Xie, Zeyuan Wang, Shuhe Huang, Haitian Liu, Ruowen Zhao, Yao Feng, Chendong Xiang, Yinze Rong, Hongyan Zhao, Hanyu Liu, Zhizhong Su, Lei Ma, Hang Su, Jun Zhu

Project

• Motus is among the earliest world-action models to unify understanding, imagination, and action through joint video-action modeling. It integrates pretrained vision-language and video-generation models with an action expert, allowing one shared model to switch flexibly across five prediction and control modes. Motion-based latent actions enable large-scale action pretraining from human videos and heterogeneous robot experience. Simulation and real-world experiments demonstrate that unifying these capabilities and pretrained priors improves downstream robot manipulation over strong contemporary baselines.

H-RDT: Human Manipulation Enhanced Bimanual Robotic Manipulation

Hongzhe Bi, Lingxuan Wu, Tianwei Lin, Hengkai Tan, Zhizhong Su, Hang Su, Jun Zhu

Project

• H-RDT is among the first studies to establish large-scale egocentric human data as a foundation for bimanual robot manipulation, addressing the scarcity of robot demonstrations. It pretrains on human videos with paired 3D hand poses, then transfers these manipulation priors to diverse robots through cross-embodiment fine-tuning. Simulation and real-world evaluations demonstrate improvements over contemporary baselines, including Pi0 and RDT, together with few-shot learning and robustness across tasks. The work has also informed the development of embodied manipulation foundation models at Horizon Robotics.

HoloBrain-0 Technical Report

Xuewu Lin, Tianwei Lin, Yun Du, Hongyu Xie, Yiwei Jin, Jiawei Li, Shijie Wu, Qingze Wang, Mengdi Li, Mengao Zhao, Ziang Li, Chaodong Huang, Hongzhe Bi, Lichao Huang, Zhizhong Su

Project

• HoloBrain-0 is a comprehensive VLA framework that bridges foundation models with reliable real-world deployment. By explicitly incorporating embodiment priors (kinematics & camera parameters), we achieve SOTA manipulation performance on both simulation and real-world benchmarks. We open-source the full ecosystem, including the RoboOrchard infrastructure, to democratize scalable robot learning.

UniMem: Towards a Unified View of Long-Context Large Language Models

Junjie Fang, Likai Tang, Hongzhe Bi, Yujia Qin, Si Sun, Zhenyu Li, Haolun Li, Yongjian Li, Xin Cong, Yankai Lin, Yukun Yan, Xiaodong Shi, Sen Song, Zhiyuan Liu, Maosong Sun

Project

• UniMem unifies long-context language modeling from the perspective of memory augmentation, turning isolated methods into a common framework for systematic analysis and design. It reformulates 16 methods across memory management, writing, reading, and injection, and analyzes representative approaches to reveal their design principles and complementary strengths. This analysis guides UniMix, which integrates those strengths to improve long-context processing and reduce perplexity relative to baselines. The work provides a systematic basis for understanding existing approaches and designing more effective memory mechanisms for long-context language models.

Beyond Fine-Tuning: Eficient and Efective Fed-Tuning for Mobile/Web Users

Bingyan Liu, Yifeng Cai, Hongzhe Bi, Ziqi Zhang, Ding Li, Yao Guo, Xiangqun Chen

Project

• This paper extend the local-user fine-tuning to multi-user fed-tuning with the help of Federated Learning (FL).