Research map / Reinforcement learning / Visual and representation RL
KungfuBot: Physics-Based Humanoid Whole-Body Control for Learning Highly-Dynamic Skills
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PBHC, a physics-based motion filtering/correction pipeline plus an adaptive tracking-factor mechanism (bi-level-derived) and asymmetric actor-critic PPO, enabling highly dynamic Kungfu/dance skills on Unitree G1.
Problem
RL motion-tracking controllers for humanoids only handle smooth, low-speed human motions because video-derived references are often physically infeasible and fixed reward tolerances fail on hard motions.
Approach
Filters and corrects video-extracted motions with physics metrics (CoM-CoP stability, contact masks), and adapts the exponential tracking-reward tolerance sigma online to the EMA of tracking error (sigma <- min(sigma, x_hat)), derived from a bi-level optimization view.
Key results
- On easy motions, PBHC reduces global MPBPE to 53.25 mm versus 233.54 mm for OmniH2O and 588.22 mm for ExBody2.
- On hard motions, PBHC achieves MPBPE 124.61 mm versus 147.88 mm (OmniH2O) and 246.40 mm (ExBody2), with lower MPJPE (1326.6 vs 1939.5).
- Real-world Tai Chi tracking on Unitree G1 closely matches MuJoCo sim-to-sim (MPBPE 36.64 mm real vs 33.18 mm sim).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PBHC instead of OmniH2O | Humanoid whole-body motion tracking · Easy | E_g-mpbpe (mm): 233.54 ± 4.013 -> 53.25 ± 17.60 | Helped |
| Use PBHC instead of OmniH2O | Humanoid whole-body motion tracking · Medium | E_g-mpbpe (mm): 433.64 ± 16.22 -> 126.48 ± 27.01 | Helped |
| Use PBHC instead of OmniH2O | Humanoid whole-body motion tracking · Hard | E_g-mpbpe (mm): 446.17 ± 12.84 -> 290.36 ± 139.1 | Helped |
| Replace adaptive tracking with one fixed tracking-factor configuration (ablation) | PBHC · Different motion types | Tracking error | Mixed |
| Use adaptive tracking instead of fixed Coarse factors (ablation) | PBHC · Jabs punch | E_g-mpbpe: 63.95 ± 6.680 -> 44.38 ± 7.118 | Helped |
| Use adaptive tracking instead of fixed Medium factors (ablation) | PBHC · Charleston dance | E_g-mpbpe: 83.63 ± 3.159 -> 94.81 ± 14.18 | Hurt |
| Use adaptive tracking instead of fixed Lowerbound factors (ablation) | PBHC · Charleston dance | E_g-mpbpe: 358.82 ± 10.35 -> 94.81 ± 14.18 | Helped |
Surprising findings
- Fixed tracking-factor settings that work well on one motion perform poorly on others, while the adaptive mechanism stays near-optimal across motions.
Limitations
- No environment awareness (terrain perception, obstacle avoidance), limiting unstructured deployment.
- Each policy imitates a single motion, not a diverse repertoire.
- Contact mask estimation reaches only 91.4% accuracy on a 10-sample manual label set.
Experimental setup
Datasets: AMASS LAFAN custom video-extracted motions
Benchmarks: custom highly-dynamic motion dataset (easy/medium/hard) MuJoCo sim-to-sim Tai Chi
Metrics: E_g-mpbpe E_mpbpe E_mpjpe E_mpbve E_mpbae E_mpjve episode length ratio
Baselines: OmniH2O ExBody2 MaskedMimic
Builds on: PPO ExBody2 OmniH2O GVHMR Reference State Initialization
Closest papers by meaning
- Universal Humanoid Motion Representations for Physics-Based Control (ICLR 2024 · similarity 0.87)
- Adversarial Locomotion and Motion Imitation for Humanoid Policy Learning (NeurIPS 2025 · similarity 0.86)
- Hierarchical World Models as Visual Whole-Body Humanoid Controllers (ICLR 2025 · similarity 0.86)
- SUGAR: A Scalable Human-Video-Driven Generalizable Humanoid Loco-Manipulation Learning Framework (NeurIPS 2026 · similarity 0.86)
- Humanoid Horizon: Extending Task Horizon in Whole-Body Loco-Manipulation via Parallel Training, Dynamic Starting, and Reward Gating (NeurIPS 2026 · similarity 0.86)
- HumanoidArena: Benchmarking Egocentric Hierarchical Whole-body Learning (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.