Reproduce “Hebbian Attractor Networks for Robot Locomotion” on OpenResearch

About this paper

Alexander Dittrich, Fuda van Diggelen, Dario Floreano

Researchers at EPFL introduced Hebbian Attractor Networks (HANs), a class of self-modifying neural networks for robot locomotion, which employ specific plasticity mechanisms to induce stable fixed-point or co-dynamic limit-cycle weight attractors. These networks achieved competitive performance against static controllers and gradient-based methods, demonstrating continuous adaptation and robustness to perturbations and morphological changes in simulated robotic tasks.

Key insights

• Max Normalization is critical for achieving stable, bounded weight dynamics and significantly improves performance in plastic neural networks for locomotion.
• The temporal structure of plasticity, specifically dual-timescale updates and activation averaging, reliably shapes emergent weight dynamics, leading to distinct fixed-point or limit-cycle attractor regimes.
• Fixed-point attractors offer robustness to external perturbations by allowing weights to reconverge to a stable configuration, while co-dynamic limit-cycle attractors are integral to the agent's oscillatory gaits and can support adaptation to morphological changes.

Summary by alphaXiv.