FAITH: Feasibility-Aware Safety-Filtered RL for High-Dimensional Systems
FAITH：兼顧可行性與高維度機器人控制的無模型安全強化學習
Traditional safe RL mixes safety with task objectives, leading to competing updates, while classical safety filters rely on exact dynamics models and suffer from short-sightedness. FAITH uses a feedforward network to approximate state-action safety values, providing a model-free filter that avoids myopic interventions. When no completely safe action exists, FAITH selects the action minimizing predicted peak harm. Tested on a 29-DoF humanoid robot and deployed on a real Unitree G1, it achieved a 99.95% safety rate while preserving 97% of unfiltered task performance.
Key points
Decoupled Task and Safety Objectives
Moves safety constraints to action filtering, preventing conflicting gradient signals in task policy updates.
Feasibility Awareness and Harm Minimization
When no safe action exists, the filter automatically chooses the action with minimum predicted peak harm, such as falling away from protected regions.
Real-World High-Dimensional Robotics
Validated on 29-DoF humanoid simulations and physical Unitree G1 hardware, achieving up to 99.95% safety rate in obstacle avoidance.
How it works
Why it matters
Balancing safety and efficiency in complex high-dimensional systems like humanoid robots has long been challenging. FAITH provides a model-free safety framework that gracefully handles infeasible states without robot collapse, marking a key step toward safe real-world deployment of humanoid robotics.
Who it affects
- AI Developer
- AI Researcher
- Student & Learner
How to use it
- 1Humanoid robot dynamic obstacle avoidance and graceful fall mitigation
- 2Safe reinforcement learning training and deployment for high-DoF robotic systems
Limitations & caveats
- Approximation errors in the safety value network may affect safety guarantees under extreme boundary conditions
- Requires training an accurate state-action safety value estimator, increasing pre-training setup complexity
Related
5 Steps to Build SimReady Robotics Assets with Frontier AI Models
5 步驟建構 SimReady 機器人資產：結合前沿 AI 模型與 NVIDIA Omniverse
A structured 5-step workflow leveraging frontier AI agents to convert, configure, and validate CAD assets into physics-ready SimReady robotics models inside NVIDIA Isaac Sim.
A Balanced Data Diet: Success Guided Sampling for Mega-Scale Robot RL
超大規模機器人強化學習的平衡數據飲食：成功引導採樣 SGS
Success Guided Sampling (SGS) focuses RL training on task configurations at the frontier of a robot's capabilities, preventing wasted compute in million-environment parallel simulations.
CSF: Contextual Safety Filtering for Text-Conditioned Motion Generators
CSF：結合場景語境的機器人動作生成安全過濾技術
CSF is a training-free framework that grounds natural-language safety rules in scene context using Control Barrier Functions to prevent unsafe robot motions.