Caltech researchers taught a Unitree G1 humanoid robot to run using a single human motion demonstration. They used mathematical optimization to turn that one demonstration into a library of different running movements for the robot to learn from. After training, the G1 reached speeds of up to 3.3 m/s... and could run outdoors while avoiding obstacles.show more

One human demo to 3.3 m/s outdoors is pretty wild. The learning curve for these robots is getting shorter fast. I keep finding examples like this that show how quickly things are moving.

We are just a radiant branch of gyroscopes

Really impressive to see how far a single demonstration can go when combined with optimization and learning. Improving data efficiency like this could be a big part of making humanoid training more scalable.

Single human demo expanded into a full running library is the same bootstrap we use for whole-body control — the optimization is doing the real work. Curious how robust 3.3 m/s stays once the terrain gets uneven.

the sim gap here is real. That single demo only works because the optimization library covers a narrow motion distribution. our domain randomization layer is built to expand that distribution so the policy doesnt collapse outdoors

That is a big step for humanoid movement learning