Hebbian Attractor Networks for Robot Locomotion

A team from EPFL, led by Alexander Dittrich, has published a groundbreaking paper on arXiv introducing Hebbian Attractor Networks (HAN). This new class of plastic neural networks is designed to give AI systems a capability largely absent in current models: the ability to continuously adapt and modify themselves in response to experiences, much like biological brains. The core innovation lies in using local Hebbian weight update normalization to induce emergent attractor dynamics. Unlike traditional static networks, HANs employ a dual-timescale plasticity mechanism combined with temporal averaging of pre- and postsynaptic neuron activations. This design allows the network's weights to converge to either stable fixed-point attractors or co-dynamic limit cycles, depending on the plasticity update frequency.

The researchers validated HANs using simulated locomotion benchmarks, gaining crucial insights into how the timing of plasticity shapes neural dynamics. Their results show that slower plasticity updates, combined with activation averaging, promote convergence to stable weight configurations ideal for consistent control. In contrast, faster updates yield oscillatory, co-dynamic systems. The team successfully generalized these findings to a high-dimensional task: controlling a simulated Unitree Go1 quadruped robot. This demonstrates that the principles of HANs can scale to complex, embodied systems, providing a principled framework for understanding the attractor regimes in self-modifying networks. The work bridges a significant gap between artificial and biological neural processing, offering a new path toward robots that can learn and adapt in real-time within unpredictable environments.