A unified task-space framework for humanoid loco-manipulation, combining multi-granularity commands, a locomotion prior with residual whole-body control, sensorless quasi-static wrist-force estimation, and downstream autonomous policy execution.

A predictive whole-body teleoperation framework that generates future robot-reference horizons from causal human observations and executes them with a generalist motion tracker. Combines flow-based retargeting, cluster-balanced training, and a residual mixture-of-experts tracker for low-latency teleoperation and autonomous loco-manipulation.

A motion-conditioned memory architecture for humanoid loco-manipulation. It uses current observations, body motion, and prior memory to select recent evidence and update a compact recurrent state, helping policies retain task-relevant information as the robot moves and its viewpoint changes.

Proposed a V-4D-A framework with Gaussian Action Field for direct action inference, showing bi-manual dexterous manipulation capability in RLBench tasks and real-world deployment. Achieved +11.54 dB PSNR improvement and boosted robotic task success rate by 10.33% over SOTA methods.

Optimized the retargeting pipeline from MoCap data to dexterous-hand interactions, designed RL reward functions for articulated object manipulation, addressed task expansion via policy distillation, and adapted motion generation to various dexterous hands.

A framework for generalizable cooperative humanoid manipulation, enabling synergy from solo-agent training to cooperative multi-agent deployment for complex manipulation tasks.