Training Physics-Based Character Controller with Residual RL and Mocap
Rudy_AA · reddit · 2026-08-14
A developer shared an experiment on building a self-balancing, physics-driven third-person character controller in Unity using reinforcement learning.
- Core Method: Built a humanoid with ArticulationBody joints and trained with Unity ML-Agents. Uses residual RL with a mocap-driven kinematic character as the reference trajectory.
- Training Details: The physics character tracks reference poses while the policy outputs residual joint corrections. Mocap provides the underlying gait, while RL learns balance, momentum control, foot placement, and recovery.
- Reward Function: Combines joint-pose imitation, end-effector tracking, COM/root velocity, orientation, height, and foot-contact agreement, with penalties for excessive actions, jitter, and abrupt changes.
More from Research
- DINOv2 Creator Introduces CAPI: A New Approach to Self-Supervised Learning — iScienceLuvr · 2026-08-14
- Extra 30% Training FLOPs Yields Inference Performance of 2x Compute — teortaxesTex · 2026-08-14
- New HF Framework: Decomposing Atomic Logic to Boost LLM Reasoning — Obed Junias · 2026-08-14
- Why Feature Engineering Needs GPU Acceleration: Iterative Search Bottlenecks — pandeyparul · 2026-08-14
- LLM Agents as Nonlinear RNNs with Exposed Hidden States — akbirthko · 2026-08-14
- DeepSeek Formalizes Testing Harness as a Research Subject — HanchungLee · 2026-08-14