Spot Nano

Walking policy trained from scratch using reinforcement learning
First known demonstration of a working RL policy for a resource-constrained quadraped (Rasberry Pi 4 CPU, no joint angle or velocity sensing)
Deployed policy to a real robot, runs at 50Hz
Approach is based on the paper "Rapid Motor Adaptation for Legged Robots" but modified to use an IMU as the only sensor input

Training and deployment of the policy. The cable is used on only for power.