# Modular Robot with AI Controller Adapts to Damage in Real Time
Researchers from Northwestern University have developed a modular robot that uses AI to generate locomotion strategies. Each module is an autonomous node with a battery, motor, processing board, proprioceptive and vestibular sensors. Links rotate 360° around the axis, enabling basic actions: jumps, turns, rolling. The structure connects at 18 points with 120° rotations, forming 435 two-module configurations. Bolt-based docking stations withstand dynamic loads.
The system relies solely on internal sensors, without any external motion capture data. The AI trains the controller on simulations that mimic evolution to optimize trajectories.
How a Single Module Works
A single unit demonstrates key maneuvers with one degree of freedom:
- Jumps (F–I): Link rotation generates impulse for lift-off.
- On-spot turns (J–M): Strategy is robust to inversion and maintains rotation direction.
- Rolling (N–P): Resists backward pushes with quick recovery.
Modules exchange data for coordination in assemblies. In case of damage or disconnection, the remaining nodes restructure movements: jumps, crawling, undulating locomotion to reach the goal.
AI Training Based on Expert Policies
The team applied an autoregressive model, similar to LLMs, to synthesize sensorimotor data. Expert controllers are trained for scenarios:
- Undamaged robot (A, F).
- Loss of one limb (B, D).
- Loss of two limbs (C).
- Loss of all four (E).
The model predicts actions from sensor history, forming an amputation-independent policy. Testing showed superiority over baseline controllers under damage:
- Removal of one rear limb: + displacement.
- Removal of both rear limbs: + displacement.
- Left with one module: + displacement.
For undamaged robots, performance matches expert level.
Demonstrations of Multi-Module Configurations
A five-module assembly masters acrobatics:
- Jumps with 66° mid-air turn (B–F).
- Self-righting from inverted position (G–L): bending and twisting to recover pose.
All configurations are trainable for these actions. The robot ignores unforeseen damage, focusing purely on getting from point A to B.
Key points:
- Modularity ensures fault tolerance: cutting it in half doesn't stop movement.
- AI controller uses only proprioceptive/vestibular data.
- Autoregressive training synthesizes policies for any damage.
- 435 two-module variants + multi-module assemblies.
- Benchmarks confirm efficiency gains with amputations.
Prospects for Robotics Applications
The approach scales to real-world tasks: search-and-rescue operations, planetary rovers. No dependence on morphology simplifies deployment in unknown environments. For mid/senior developers, the key is transfer learning from expert policies to generative models without motion capture. Implementation requires simulators with hinge physics and high-load connections.
— Editorial Team
No comments yet.