Excited to share our latest work: “Bio-Inspired Plastic Neural Networks for Zero-Shot Out-of-Distribution Generalization in Complex Animal-Inspired Robots” 🪲🦎 We show that Hebbian learning outperforms LSTM-based adaptation for real-world transfer. It even works without domain randomization! It can handle: ✅ Uneven terrain ✅ Morphological damage ✅ Sim-to-real gaps

Current ML methods are powerful but brittle. Most fail when conditions change — sim-to-real gaps, uneven terrain, or even minor morphological damage can cause catastrophic failures. Domain randomization is a way to increase robustness but can be expensive to train.

Inspired by biological brains, we use Hebbian plasticity where synaptic weights adapt in real-time based on neural activity. This allows dynamic adaptation, even in unseen environments and seems to not crucially depend on domain randomization. We tested this approach on two challenging, high-DOF robots: 🪲 An 18-DOF dung beetle-inspired robot 🦎 A 16-DOF gecko-like robot Both showed zero-shot sim-to-real transfer, adapting instantly without retraining.

What type of solution did the Hebbian network find? PCA of weight dynamics revealed limit cycle attractors, suggesting rhythmic motor control emerges from plasticity + sensory feedback. A limit cycle attractor can be observed in the weights of the ANN with Hebbian updates (left). In contrast, non-optimized Hebbian rules cannot achieve a limit cycle attractor and the weights were continually updated toward a single direction in the weight space (right).

We believe this work highlights the power of bio-inspired learning mechanisms in robotics. Our Hebbian networks are simple, lightweight, and robust and they generalize without requiring architectural complexity or training tricks. Paper: Project page with videos: A great collaboration with @BinggwongLeung , @worasuchad, @JoachimWinther. and @PManoonpong.

really cool. has hebbian learning been explored in combination with topology evolving networks, e.g. NEAT, or is the parameter search space on a fixed architecture already difficult to optimize?

It has actually! Here is a review article of that field: Although the article is now a few years old so maybe time for an update.

Zero shot is terrifying

Great work!!!

@drmichaellevin Grow up from your toys If you ready 🙊

Interesting!

Is there a reason why these bio inspired approaches don’t get applied to language? It’s always robots automata etc ..

Compared to SNN ?

From semiconductors to data centers, AIS targets the critical components behind AI's exponential growth. Capture potential returns from this transformative technology sector.