Robot learning · Embodied intelligence
I am a master’s student in Electronic and Information Engineering at the University of Science and Technology of China (USTC). I work on robot learning and embodied intelligence, with an emphasis on making manipulation policies adapt to physical interaction and work reliably on real robots.
My work covers the full experimental pipeline: robot and camera calibration, data collection, simulation, model training, inference, control interfaces, and real-robot evaluation. At the Institute for AI Industry Research (AIR), Tsinghua University, I worked with robot arms and a dexterous hand on the Zeva, Zeva-Ego, and Zetta projects.
Zeva encodes executed actions and observed state changes as causal interaction signals. A dual-timescale memory retrieves useful experience to adapt a frozen policy without gradient updates. It reached 76.8% average success on RoboCasa365-Atomic5; in a separate repeated-attempt evaluation, cumulative success rose from 26% on the first attempt to 73% within four attempts.
Zeva-Ego turns egocentric video into action-centered supervision for VLA mid-training, then uses interaction feedback for parameter-free adaptation at deployment. Scaling to 10,000 hours of video improved RoboTwin success from 63.8% to 75.3%; repeated-attempt success rose from 58% to 89%.
Zetta keeps a base policy frozen while runtime critics detect failures and recovery skills support continued execution. Rollout analysis and validation update those skills over time. The paper reports 90.8% success on LIBERO-Pro, 93.6% on RoboCasa, and an 11.1× inference speedup. I worked on the base policy and implemented real-robot deployment.
* Equal contribution, as indicated in the papers.
Worked on VLA and world-action models for robot manipulation across data collection, training, inference, and real-robot evaluation. Integrated robot arms and a dexterous hand, including calibration and control interfaces. Built a Real2Sim2Real workflow for simulation rollouts, policy training, and transfer back to physical robots.
Ran Franka manipulation experiments and reproduced Seer, a predictive inverse-dynamics model, including its training and inference pipeline.
Developed a C++ SDK for robot walking speed and mode switching. Adapted ModelBasedFootstepPlanning to a 12-DoF biped, trained in simulation, and validated sim-to-sim and sim-to-real deployment with MuJoCo and RKNN.
M.Eng., Electronic and Information Engineering
B.Eng., Electronic and Information Engineering
Led a modular embedded control framework with 400+ stars, recognized with the RoboMaster 2023 Open Source Award (third prize). Developed a wheel-legged balancing infantry control system using Kalman filtering and LQR; the team won first prize in the 2023 RoboMaster University League infantry robot competition.
Programming: Python, C++ · Robot systems: ROS/ROS 2, MuJoCo, Isaac Sim, RKNN · Machine learning: PyTorch, Transformers, distributed training, VLA and world-action models