Low-Cost Bio-Inspired Controllers Outperform Overparametrized AI in Robot Learning
A new arXiv preprint (2604.20365) challenges the trend toward overparametrization in robotics, showing that simpler bio-inspired controllers can outperform larger neural networks. Researchers compared Central Pattern Generators (CPGs) and Multi-Layer Perceptrons (MLPs) in optimizing robot control for a given morphology with limited proprioception. Using evolutionary and reinforcement learning protocols, they varied parameter spaces across multiple reward functions. Results showed that shallow MLPs and densely connected CPGs achieved better performance than deeper MLPs or Actor-Critic architectures. The study introduces a metric, Paramete, to account for the relationship between performance and parameter count. The findings suggest that in contexts with small input/output spaces and bounded performance, more parameters can hinder learning. This empirical work provides evidence for the benefits of low-cost bio-inspiration in an era of overparametrization, with implications for efficient robot control design.
Key facts
- arXiv preprint 2604.20365 compares CPGs and MLPs in robot control optimization
- Study uses a robot morphology with limited proprioceptive capabilities
- Evolutionary and reinforcement learning protocols were applied
- Shallow MLPs and densely connected CPGs outperformed deeper MLPs and Actor-Critic architectures
- Multiple reward functions were used to vary parameter spaces
- Overparametrization can hinder learning in small input/output spaces
- A new metric called Paramete is introduced
- Research was submitted to arXiv on April 26, 2025
Entities
Institutions
- arXiv