Chinese scientists are making robots think instantly, accelerating the development of silicon photonics

News about a breakthrough in photonic neuromorphic computing

Scientists from Xidian University have created the first fully optical “neuromorphic” chip capable of reinforcement learning without converting signals into electrical current. This event marks the transition from linear photonic spiking networks to nonlinear transformations—a key step for practical applications.

Why it matters
- No conversion: Translating photons to electrons and back causes energy and time losses. In real‑time systems (robotics, autopilots) such delays can lead to equipment failure or even accidents.
- Safe human–robot interaction: Developing universal photonic chips paves the way for more reliable and energy‑efficient intelligent systems.

Three solved problems
1. Availability of large arrays of nonlinear spiking neurons with low activation thresholds—now neurons can be packed more densely.
2. Fully programmable chips—previously they were “hard” (hardware‑programmed).
3. Photonic reinforcement learning—now realized thanks to a new architecture.

Prototype architecture
Component | Description
---|---
16‑channel photonic chip | Contains 272 trainable parameters, built on a 16×16 Mach–Zehnder interferometer matrix.
Chip with lasers and feedback | Uses a saturable absorber for low‑threshold nonlinear spike activation.
Hardware‑software framework | First trained in software, then ported to chips, after which fine‑tuned considering hardware specifics.

Testing
- CartPole (pole balancing) – accuracy almost identical to the software model (1.5 % drop).
- Pendulum (inverted pendulum) – 2 % drop.
- On both tasks, computation latency was only 320 ps.

Efficiency
Type | Energy consumption | Density
---|---|---
Linear | 1.39 TOPS/W | 0.13 TOPS/mm²
Nonlinear | 987.65 GOPS/W | 533.33 GOPS/mm²

These figures place the photonic system in the GPU class for energy efficiency (~1 TOPS/W) and density (0.1–0.5 TOPS/mm²), while relying entirely on optical processing, eliminating conversion losses.

Prospects
- Autonomous driving
- Intelligent robots
- Edge computing with ultra‑low latency and minimal power consumption

Future plans include scaling the chip to 128 channels to tackle more complex tasks (neuromorphic autonomous navigation) and creating compact hybrid‑integrated photonic neuromorphic devices.

In summary: The development opens a new path to energy‑efficient AI based on light pulses, which could radically change approaches to robotics and autonomous systems.