A laser-based AI navigator used little energy for signals faster than a billionth of a second

Laser signals for navigation

The proposed system uses a light-based spiking Actor network—the part of the reinforcement-learning system that chooses actions—alongside two continuous-valued Critic networks that evaluate those actions. The Actor’s final nonlinear spiking activation layer was implemented with a distributed-feedback laser array containing saturable absorbers.

In autonomous navigation tasks, the system achieved an average reward of 58.22 ± 17.29 and a success rate of 80% ± 8.3%. Hardware-software co-inference estimated energy use of 0.78 nJ per device-level nonlinear activation and a latency of 191.20 ps under ideal parallel channel activation. The reported co-inference error rates were 0.051% with obstacle interference and 0.059% without it. Simulations of channels activated by errors matched the expected responses, supporting the reported dynamic behavior of the laser devices.

Why the hardware matters

Conventional electronic computers can face energy and delay costs when running reinforcement-learning systems. The reported results indicate that a photonic, or light-based, implementation could provide low-energy nonlinear activations with very short latency for an autonomous-navigation architecture. The findings point to a possible route for reducing those costs, while the study’s partial hardware validation leaves the performance of a complete physical navigation system unresolved.

Evidence and caveats

This is a journal article reporting simulations of autonomous navigation and hardware-software co-inference using an array of distributed-feedback lasers with saturable absorbers. The hardware test covered the Actor network’s final nonlinear activation layer rather than the complete navigation system. The reported latency assumes ideal parallel activation of channels, and the abstract does not establish performance in a fully physical, moving robot.