Triple-Gate Actuation: A Hardware-Isolated Safety Architecture for Adversarial Human-Robot Physical Interaction

摘要

We present a theoretical three-layer isolation architecture for robotic systems that deliberately apply physical force to human users — specifically, adaptive training systems that provide continuous, individualized, escalating resistance while maintaining safety through hardware-enforced constraints designed to be architecturally independent of the AI behavior system. The architecture addresses a gap in existing human-robot safety frameworks, which assume cooperative interaction: the proposed system must be safe while intentionally applying force in an adversarial training context. The architecture's core safety property is that the AI model interacts with the physical world exclusively through an abstract function interface, all commands pass through a mandatory hardware validation gate before reaching the motor drivers, and the motor drivers themselves require three independent, concurrent inputs to produce output: a validated command, a sensor-derived power budget signal, and a real-time consistency check between the budget signal and the safety layer's internal accumulator. No single gate bypass is sufficient; the three gates form a mutually reinforcing triad in which each gate's failure mode is caught by the others. Maximum system output is physically constrained to never exceed what the safety layer can arrest mid-motion, making the system designed to be safe by construction rather than by software correctness. The framework is applicable beyond combat training to domains including surgical training, physical rehabilitation, ergonomic injury prevention, and athletic development, where the core requirements — precisely calibrated adaptive resistance, hardware-enforced safety limits, and structured data capture — are shared. No prototype has been built; the architecture is a conceptual design intended to establish theoretical foundations and identify key engineering constraints for future implementation.