Battle Angel

A Unitree G1 humanoid that walks into collapsed buildings to find survivors. The robot learns to move through rubble and hazards using reinforcement learning in MuJoCo, staying upright and dodging debris as it goes. Above it, Gemini acts as the brain: given a rescue mission in plain language, it decides which survivor to reach first, issues waypoints, watches the results come back, and adapts its plan when a run fails. Gemini reasons about what to do, the learned policy handles how to move. A FastAPI backend runs rescues on demand and reports live metrics for every attempt, survivor reached, distance, falls, hazard steps, and gait quality, with full trajectories you can replay. A glimpse of robots that pair learned movement with real reasoning to go where people cannot.