Japanese researchers train robot swarms using single-robot imitation learning
Researchers demonstrated that up to 40 Raspberry Pi-based robots can learn coordination behaviors from demonstrations of a single robot, without inter-robot communication or external positioning.
Summary
A team from Japan's National Institute of Advanced Industrial Science and Technology, Institute of Science Tokyo, and University of Tokyo developed an imitation learning approach for multi-robot coordination. Using a lightweight 2.4-million-parameter neural network running on Raspberry Pi 4 hardware, groups of wheeled robots learned to collect and push blocks in a shared workspace. The robots operate independently with only forward-facing cameras, processing visual input at 5 Hz without GPU acceleration.
How the swarm learns from one
- Hardware platform: Raspberry Pi Mouse V3, 14 cm wide, with Raspberry Pi 4 and Camera Module 3
- Neural network: MobileNetV2 for visual processing, mixture-of-experts with two LSTM-based expert networks and a routing network
- Model size: 2.4 million parameters, runs on CPU without GPU acceleration
- Control rate: 5 Hz velocity command generation from camera observations
- Training data: Human joystick demonstrations of single robot, with two conditions—stationary background robots or autonomously moving background robots
- Scale tested: Up to 40 robots in 3.5-meter-square workspace with 28 orange blocks
Why it matters
The approach lowers the data collection barrier for swarm robotics by eliminating the need for multi-robot teleoperation, though congestion limits and the difficulty of learning from moving-robot demonstrations indicate significant scaling challenges remain before practical deployment.
Original reference: www.hackster.io →