Thrilled to share our #SIGGRAPHAsia2025 work, PhysHMR: Learning Humanoid Control Policies from Vision for Physically Plausible Human Motion Reconstruction (fengq1a0.github.io/projects/…).

✨ Highlights: • End-to-end visual→action (no multi-stage pipelines, less error accumulation) • Efficient training via supervised distillation RL (PPO) • Real-world transfer: the policy can also drive a Unitree G1 robot 🤖 Check the videos 👇