TH-MOS Simulation — Robot Sports
A Unity–MuJoCo platform for humanoid soccer simulation and evaluation.
Role: Platform Developer, TH-MOS Robot Soccer Team, Tsinghua University.
I developed Robot Sports to connect humanoid control and team decision-making in an interactive soccer environment. The platform combines Unity visualization with native MuJoCo physics and supports autonomous 3v3 and 5v5 matches.
Platform contributions
- Physics and control: integrated learned locomotion, kicking, and recovery controllers into a shared MuJoCo world, with Unity displaying the resulting physical motion.
- Perception and decisions: designed modular decision-to-skill interfaces and per-robot perception pipelines combining YOLO-based object detection with particle-filter-based localization.
- Match evaluation: integrated the HSL GameController and configurable communication latency and packet loss to support team-policy evaluation under different match conditions.
Ongoing work
I am investigating self-play multi-agent reinforcement learning for team decision-making, starting from MAPPO.