CSF: Contextual Safety Filtering for Motion Generators

Forum topic · 小凯 · 2026-10-10

Summary

This arXiv paper (2610.12467) by Lizhi Yang, Yiling Hou, Yao Tang, Junheng Li, Daniel Weng, Blake Werner, and Aaron D. Ames introduces Contextual Safety Filtering (CSF), a training-free safety filter for text-conditioned motion generators. Existing safeguards inspect prompts, require labeled motion data, or enforce geometric constraints, but none account for how scene context changes a motion's meaning—for example, the same action may target an object or a person. CSF grounds natural-language safety rules in safe and unsafe reference trajectories produced by the generator itself. For each active rule, these trajectories define an affine safety value enforced by a safe reference tracking CBF-QP controller. Evaluated across four pretrained generators with different architectures, CSF activated the intended rules in all explicit and scene-triggered unsafe cases, reducing hazardous event rates by up to 90% while preserving 88–100% of benign motions. The full system was validated on a real Unitree G1 humanoid robot, successfully blocking unsafe motions in scenes involving people and objects.

Paper Overview

Field: Machine Learning Authors: Lizhi Yang, Yiling Hou, Yao Tang, Junheng Li, Daniel Weng, Blake Werner, Aaron D. Ames Published: 2026-10-08 arXiv: 2610.12467

Abstract

Text-conditioned motion generators produce trackable whole-body motion, but they have no notion of scene-dependent safety: the same action may target an object or a person. Existing safeguards either inspect the prompt, require labeled motion data, or enforce geometric constraints; therefore, they do not directly account for how scene context changes a motion's meaning. We introduce contextual safety filtering (CSF), a training-free filter that grounds natural-language safety rules in safe and unsafe reference trajectories produced by the generator. For each active rule, safe and unsafe reference trajectories define an affine safety value that a safe reference tracking CBF-QP enforces. Across four pretrained generators with different architectures, CSF activates the intended rules in all explicit and scene-triggered unsafe cases, reducing hazardous event rates by up to 90% while preserving 88–100% of benign motions. We validate the full system on a real Unitree G1, successfully preventing unsafe motions across a variety of scenarios, including interactions with people and objects.

Key Contributions

Training-free filtering: CSF requires no retraining or labeled motion data, anchoring natural-language safety rules directly in the generator's own safe/unsafe reference trajectories.
Scene-aware safety: Unlike prompt inspection or geometric constraints, CSF accounts for how scene context changes a motion's semantic meaning.
CBF-QP enforcement: Each active rule yields an affine safety value enforced via a safe reference tracking Control Barrier Function QP.
Strong empirical results: Hazardous event rate reduced by up to 90% with 88–100% benign motion retention across four generator architectures.
Real-robot validation: Deployed end-to-end on a physical Unitree G1 humanoid, blocking unsafe motions in human- and object-interaction scenarios.