Workhorse teaches a Unitree G1 whole-body skills from human video
The G1 topples and climbs a suitcase, catches a thrown box and sorts boxes with a kick, and Liao says the clip is one take.
Songbo Hu and Qiayuan Liao have released Workhorse, a method for learning whole-body loco-manipulation on a humanoid directly from human demonstrations. No robot teleop (a human driving the robot remotely). No retargeting (mapping a human's joint motions onto the robot's different body).
Can a humanoid learn robust whole-body loco-manipulation directly from a human? No robot, no teleop, no retargeting.
On a Unitree G1, the result is a short list of things humanoids usually do badly. The robot topples a suitcase, climbs on it, catches a thrown box, and sorts boxes using both its hands and a kick. Hu says it is fully autonomous.
Why the kick matters
The interesting part is that the skills are not arm skills. Toppling a suitcase and then standing on it means the whole body is the manipulator, legs included, and the robot has to stay upright while the thing it is pushing moves. Catching a thrown box is a timing problem the robot cannot slow down to solve. Those are the cases where most manipulation stacks, which assume a stable base and a gripper at the end of an arm, run out of road.
The pitch is that all of this came from watching a human, not from someone wearing a VR rig and driving the G1 through hundreds of attempts. Teleop data collection is the expensive, slow part of the current humanoid playbook, and retargeting is the step where human motion gets mangled into robot motion and quietly loses whatever made it work.
Shown versus claimed
Liao posted his favourite clip and described it as one take, real time and fully autonomous, and said the team tried everything to make it fail. That claim is what separates this from a cut-together highlight reel, and it is the team's own claim. Nobody outside the project has run the robot yet.
More videos are on the project page linked from Hu's post. The sources do not say whether code or a paper is out.
we tried everything to make it fail🏒