Living brain cells outperform silicon AI in embodied navigation task

A team of 15 researchers from institutions including Cortical Labs, Honda Research Institute, and universities has published a paper detailing the Embodied Neurocomputation framework, which connects living biological neural cultures (BNNs) to traditional silicon computing for task-driven learning. In their experiments, they placed BNN agents in a simulated grid-world where they had to navigate an odor-style gradient via closed-loop feedback. The team systematically optimized encoding and decoding parameters across a massive combinatorial space, evaluating roughly 1,300 parameter combinations over more than 4,000 hours of real-time agent-environment interactions. They identified 12 configurations that demonstrated consistent learning across multiple episodes, achieving significantly higher task performance than optimized silicon-based Deep Q-Network (DQN) agents under the same interaction budget.

This work marks a major step toward scalable goal-oriented learning using living neural tissue. The framework establishes a foundation for task-driven neurocomputing benchmarks and supports the development of hybrid bio-silicon architectures. Potential applications include ultra-energy-efficient robotic control, where biological neurons' natural adaptability and low power consumption could complement silicon's speed. The researchers emphasize that even simple biological interactions gave rise to rich, multi-combinatorial search spaces, hinting at the untapped computational capacity of living neurons. Future work will focus on scaling the task complexity and refining interfaces for real-time, adaptive computation in physical robots.