CSF: Contextual Safety Filtering for Motion Generators
The paper proposes a training-free safety filter that uses scene context to stop unsafe robot motions before execution.
CSF grounds natural-language safety rules in safe and unsafe reference trajectories generated by the same motion model. It then enforces those rules with a reference-tracking CBF-QP, so the same motion can be treated differently depending on whether it targets an object or a person. Across four pretrained generators, the authors report intended rule activation in explicit and scene-triggered unsafe cases, danger-event reductions up to 90%, and 88-100% preservation of benign motions. They also demonstrate the system on a Unitree G1 in human- and object-interaction scenarios. ArXiv · AI/CL/LG's note
CSF grounds natural-language safety rules in safe and unsafe reference trajectories generated by the same motion model. It then enforces those rules with a reference-tracking CBF-QP, so the same motion can be treated differently depending on whether it targets an object or a person. Across four pretrained generators, the authors report intended rule activation in explicit and scene-triggered unsafe cases, danger-event reductions up to 90%, and 88-100% preservation of benign motions. They also demonstrate the system on a Unitree G1 in human- and object-interaction scenarios. ArXiv · AI/CL/LG's note
score 4