Description

Soft robotic arms are hard to model in real-time. The standard continuum mechanics equations are accurate but computationally expensive—too slow for closed-loop control. This project replaces those calculations with a recurrent neural network that runs fast enough for real-time use.

The key insight: separate the position and orientation equations. This decomposition makes the problem tractable for neural networks. I used Long Short-Term Memory (LSTM) layers because they handle sequential data well and retain information across longer time horizons without gradient issues.

Unlike classical models that ignore friction, this approach accounts for it. The network learns the full dynamics including contact forces and damping.

Testing used a magnetically-controlled soft robotic arm. The RNN model predicted the arm’s motion more accurately than a classical Cosserat rod model, and ran fast enough to close the control loop.

The result is a lightweight model that captures soft robot dynamics at speeds suitable for real-time control.