Embodied physical learning: a trained bacterial cell wall can guide organism behavior
Abstract
Physical learning studies how physical systems themselves can be trained to perform learning tasks using physical laws. For instance, disordered spring networks can be trained by local rules to route a mechanical signal from one site in the network to another. Disordered spring networks play key roles in many facets of cell biology, not just in terms of mechanics, but also in terms of motility. How then can training on the substrate itself impact the action of an organism, such as a bacterium? Here, we train a spring-network model of the bacterial cell wall and embed it in a motile agent. The stimulus is the load the cell generates by colliding with a surface. The trained wall routes that load across the body to distant flagellar motors, where it drives stator release, and the resulting turn biases the cell toward escape and delays subsequent surface contacts. On the first near-wall tumble after entrapment, the trained cell recovers 89% of the advantage an ideal side-resolved controller provides. We call this arrangement embodied physical learning: the stimulus, the actuator, and the consequences all belong to the agent, and the training force scale is inherited from the agent's own sensing rather than imposed. An actuator with its own kinetics imposes a second, tighter constraint: the routed signal must switch the actuator within the stimulus window. We also find that behavior converges within a few hundred wall encounters, comparable to what a cell in sustained contact with structure experiences in a single generation, and that the routing survives continuous crosslink remodeling provided crucial, tension-bearing bonds are protected. Experimental tests of embodied physical learning are discussed.
Related
- Predicting Locomotor Behavior on an Irregular Surface from Even-Surface Gait: A Machine Learning Approach
- Sensor-aware physical modelling for deep learning-based single-photon imaging
- On physical implementability of reinforcement-learning-driven reasoning
- A Machine-Learning Based Approach to the Evaluation of the Critical Scaling Behavior of Anisotropic Spin Systems
- Knowledge-Guided Construction of Machine Learning Potentials for Reactive Behavior in Energetic Materials
BibTeX
@misc{elmeligy2026embodied,
title = {{Embodied physical learning: a trained bacterial cell wall can guide organism behavior}},
author = {Nada Elmeligy and No Chen and J. M. Schwarz},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2611.07112},
eprint = {2611.07112},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.07112},
}
RIS
TY - UNPB
TI - Embodied physical learning: a trained bacterial cell wall can guide organism behavior
AU - Nada Elmeligy
AU - No Chen
AU - J. M. Schwarz
PY - 2026
DA - 2026/11/04
PB - arXiv
DO - 10.48550/arxiv.2611.07112
UR - https://arxiv.org/abs/2611.07112
AB - Physical learning studies how physical systems themselves can be trained to perform learning tasks using physical laws. For instance, disordered spring networks can be trained by local rules to route a mechanical signal from one site in the network to another. Disordered spring networks play key roles in many facets of cell biology, not just in terms of mechanics, but also in terms of motility. How then can training on the substrate itself impact the action of an organism, such as a bacterium? Here, we train a spring-network model of the bacterial cell wall and embed it in a motile agent. The stimulus is the load the cell generates by colliding with a surface. The trained wall routes that load across the body to distant flagellar motors, where it drives stator release, and the resulting turn biases the cell toward escape and delays subsequent surface contacts. On the first near-wall tumble after entrapment, the trained cell recovers 89% of the advantage an ideal side-resolved controller provides. We call this arrangement embodied physical learning: the stimulus, the actuator, and the consequences all belong to the agent, and the training force scale is inherited from the agent's own sensing rather than imposed. An actuator with its own kinetics imposes a second, tighter constraint: the routed signal must switch the actuator within the stimulus window. We also find that behavior converges within a few hundred wall encounters, comparable to what a cell in sustained contact with structure experiences in a single generation, and that the routing survives continuous crosslink remodeling provided crucial, tension-bearing bonds are protected. Experimental tests of embodied physical learning are discussed.
ER -