Piers Millar

Unitree G1 just learned how to rescue itself when trapped under obstacles. Chinese researchers developed ROOF, a reinforcement learning strategy that teaches G1 to figure out how to escape. Instead of relying on cameras or pre-built 3D maps, the system uses a geometric reward to guide the G1 as it organically senses spatial constraints. The less the robot overlaps with obstacles and the greater the distance it creates, the more reward it receives. But the researchers don't prescribe how or which direction it should move. G1 has to figure that out on its own. With this guidance, G1 learns to shift its body, use its limbs to find stable support, push movable obstacles aside, and stand up using proprioception alone. In real-world tests, the G1 successfully escaped and stood up in 95 out of 150 trials across 7 trapped environments, a 63.3% success rate. ROOF turns getting up while trapped into a skill that robots can learn and measure.

I am witnessing the Watchhousification of burger concepts