Researcher Trains 14 Robot Skills in 5 Days Using One GPU and Queued RL

A developer trained a bipedal robot duck called MicroDuck to perform 14 distinct skills — including walking, skating, kicking, and object picking — using reinforcement learning on a single GPU over approximately five days. The robot uses 15 servos and runs on a 50Hz control loop, with all skills trained via PPO across up to 4,096 parallel simulation environments. Tasks were queued serially to avoid GPU memory conflicts, and completed runs were marked with a checkpoint file to enable automatic retries on failure. A key challenge was out-of-memory crashes on rough-terrain tasks, which were resolved by halving the environment count to 2,048. All 14 trained policies were exported in ONNX format for on-device deployment, with rollout animations used as the primary acceptance test rather than loss curves alone.