Two October 2026 papers filter robot actions on Unitree G1

FAITH and CSF, posted 8 October 2026, filter robot actions to cut safety violations while keeping most task or motion performance.

On 8 October 2026, two arXiv papers described separate robot safety filters, both demonstrated on a Unitree G1. FAITH is a model-free safe reinforcement-learning method that approximates an optimal state-action safety value and amortizes minimal-intervention filtering with a feedforward network; the task policy optimizes return through the filtered dynamics, and when no safe action exists the filter approaches the action with minimum predicted peak harm. On a double integrator and Safety Gym it reported the highest return among methods with no feasible-start violations, and on a 29-DoF humanoid Walking-Avoid task it reported a 99.95% safety rate while retaining 97% of unfiltered return. Separately, Contextual Safety Filtering is training-free: it grounds natural-language safety rules in safe and unsafe reference trajectories from a text-conditioned motion generator and enforces an affine safety value with a control-barrier-function quadratic program. Across four pretrained generators it activated the intended rules in all explicit and scene-triggered unsafe cases, reduced the danger-event rate by up to 90%, and preserved 88-100% of benign motions, and a Unitree G1 system blocked unsafe motions involving people and objects. The sources do not disagree; they cover different methods and evaluation metrics.

• FAITH (arXiv, 2026-10-08): model-free safe RL that learns a state-action safety value and a feedforward minimal-intervention filter; the task policy optimizes return through the filtered dynamics.
• When no safe action exists, FAITH approaches the action with minimum predicted peak harm.
• On a double integrator and Safety Gym, FAITH reported the highest return among methods with no feasible-start safety violations.
• On a 29-DoF humanoid Walking-Avoid task, FAITH reported a 99.95% safety rate and 97% of unfiltered return; policies were shown on a Unitree G1.
• CSF (arXiv, 2026-10-08): training-free Contextual Safety Filtering grounds natural-language rules in safe and unsafe trajectories from a text-conditioned motion generator and enforces an affine safety value with a control-barrier-function…
• Across four pretrained generators, CSF activated the intended rules in all explicit and scene-triggered unsafe cases, cut the danger-event rate by up to 90%, and preserved 88-100% of benign motions.
• A complete CSF system on a Unitree G1 blocked unsafe motions involving people and objects.
• The reports describe different methods and metrics and do not contradict each other.