SAGE: A Safety-Aligned Gradient Enforcement Method for Human-Machine Collaboration

SAGE addresses two types of mismatch between CBF filtering and multi-agent reinforcement learning in human-machine collaboration, achieving a 71.0% success rate and only 0.5 collision steps per thousand environment steps across nine simulated scenarios. Its SAIL internalizes safety constraints while execution still uses CBF quadratic programming, and TALO updates independently with a Lyapunov-corrected specification. Ablations show that direct CBF filtering reduces collision frequency by 98.5% but yields only a 59.3% success rate; experiments with two humanoid robots verified deployment feasibility.