Research map / Adversarial robustness and security / Attacks on text and graphs
Keep On Going: Learning Robust Humanoid Motion Skills via Selective Adversarial Training
Yang Zhang, Zhanxiang Cao, Buqing Nie, Haoyang Li, Zhong Jiangwei, Qiao Sun, Xiaoyi Hu, Xiaokang Yang, Yue Gao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace domain randomization with selective adversarial training | perceptive locomotion policy · Perceptive locomotion, clean env (sim) | velocity tracking error: 0.25 -> 0.19 | Improved |
| replace domain randomization with selective adversarial training | perceptive locomotion policy · Perceptive locomotion, DR env (sim) | velocity tracking error: 0.28 -> 0.21 | Improved |
| replace domain randomization with selective adversarial training | perceptive locomotion policy · Perceptive locomotion, DR env (sim) | gravity projection error: 0.42 -> 0.36 | Improved |
| replace domain randomization with selective adversarial training | perceptive locomotion policy (real G1) · Real-robot traversal: Stairs | traversal success rate (%): 47 -> 84 | Improved |
| replace domain randomization with selective adversarial training | perceptive locomotion policy (real G1) · Real-robot traversal: Slopes | traversal success rate (%): 54 -> 90 | Improved |
| replace domain randomization with selective adversarial training | whole-body control policy (real G1) · Real-robot dance imitation | trajectory tracking error reduction (%): 32% | Improved |
| replace domain randomization with selective adversarial training | whole-body control policy (real G1) · Real-robot dance imitation (120 s) | duration before termination (s): 65 -> 120 | Improved |
| use selective attack policy instead of always-on (full-time) attack policy (ablation) | adversarial training · Perceptive locomotion (sim), perturbation levels 1-4 | normalized reward | Improved |
| use selective attack policy instead of always-on attack policy (ablation) | adversarial training · Perceptive locomotion, unperturbed env | normalized reward | Improved |
| use full-time attack policy (PAP) instead of domain randomization (ablation) | adversarial training · Perceptive locomotion, low-perturbation env | normalized reward | Hurt |
| increase attack-budget penalty lambda from 10^-3 to 10^-1 (ablation) | SAP (WBC) · Whole-body control (sim) | reward | Improved |
| increase attack-budget penalty lambda from 10^-1 to 10^1 (ablation) | SAP (WBC) · Whole-body control (sim) | reward | Hurt |
Closest papers by meaning
- Adversarial Locomotion and Motion Imitation for Humanoid Policy Learning (NeurIPS 2025 · similarity 0.89)
- Humanoid Horizon: Extending Task Horizon in Whole-Body Loco-Manipulation via Parallel Training, Dynamic Starting, and Reward Gating (NeurIPS 2026 · similarity 0.85)
- Universal Humanoid Motion Representations for Physics-Based Control (ICLR 2024 · similarity 0.85)
- Enhancing Robustness in Deep Reinforcement Learning: A Lyapunov Exponent Approach (NeurIPS 2024 · similarity 0.84)
- Hybrid Internal Model: Learning Agile Legged Locomotion with Simulated Robot Response (ICLR 2024 · similarity 0.84)
- Hierarchical World Models as Visual Whole-Body Humanoid Controllers (ICLR 2025 · similarity 0.84)
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.
