Bio-Inspired Plastic Neural Networks for Zero-Shot Out-of-Distribution Generalization in Complex Animal-Inspired Robots

Binggwong Leung

Worasuchad Haomachai

Joachim Winther Pedersen

Sebstian Risi

Poramate Manoonpong

Vidyasirimedhi Institute of Science & Technology, Rayong, Thailand
IT University of Copenhagen, Copenhagen, Denmark
University of Southern Denmark, Odense, Denmark

Accepted at IROS 2025.

Abstract

Artificial neural networks can be used to solve a variety of robotic tasks. However, they risk failing catastrophically when faced with out-of-distribution (OOD) situations. Several approaches have employed a type of synaptic plasticity known as Hebbian learning that can dynamically adjust weights based on local neural activities. Research has shown that synaptic plasticity can make policies more robust and help them adapt to unforeseen changes in the environment. However, networks augmented with Hebbian learning can lead to weight divergence, resulting in network instability. Furthermore, such Hebbian networks have not yet been applied to solve legged locomotion in complex real robots with many degrees of freedom. In this work, we improve the Hebbian network with a weight normalization mechanism for preventing weight divergence, analyze the principal components of the Hebbian’s weights, and perform a thorough evaluation of network performance in locomotion control for real 18-DOF dung beetle-like and 16-DOF gecko-like robots. We find that the Hebbian-based plastic network can execute zero-shot sim-to-real adaptation locomotion and generalize to unseen conditions, such as uneven terrain and morphological damage.

Experiments

The initial weight configuration for the control network

The weights for the FF network are initialized uniformly with $w\in(-0.1,0.1)$ .
The weights of the Hebbian network are initialized uniformly with $w\in(-0.01,0.01)$ , whereas the Hebbian coefficients are initialized with a normal distribution with a mean of zero and a standard deviation of 0.01.
The weights of the LSTM network are initialized with $w\in(-0.1,0.1)$ , whereas the hidden and cell states are initialized with $w\in(-0.01,0.01)$ .

These initializations were found through hand-tuning in preliminary experiments.

Citation:

@misc{leung2025bioinspiredplasticneuralnetworks,
      title={Bio-Inspired Plastic Neural Networks for Zero-Shot Out-of-Distribution Generalization in Complex Animal-Inspired Robots}, 
      author={Binggwong Leung and Worasuchad Haomachai and Joachim Winther Pedersen and Sebastian Risi and Poramate Manoonpong},
      year={2025},
      eprint={2503.12406},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2503.12406}, 
}

If you have any questions or doubts about this project, you are welcome to contact me. My email address is haomachai@gmail.com