RAI Institute’s system learns from human demonstrations, then performs selected behaviors on a real Unitree G1.
IEEE Spectrum featured Workhorse, a RAI Institute system shown performing contact-rich whole-body manipulation. The demonstration stage matters: this is a video showing selected tasks, not evidence of a general-purpose robot.
Workhorse’s central idea is to learn from human demonstrations collected without the robot performing the demonstration itself. That could reduce the need to make a humanoid repeatedly attempt difficult actions while collecting its own training data.
“Contact-rich” tasks involve several changing points of contact, such as hands, feet, and an object. RAI Institute says humanoids still struggle to plan these behaviors using cameras mounted from the robot’s viewpoint and proprioception, its sense of joint positions and movement.
On a real Unitree G1, Workhorse sorts boxes with its hands and a kick, catches a thrown box, then topples and climbs a suitcase, according to IEEE Spectrum. These examples show coordinated movement across the robot’s body rather than simple arm motion.
The supplied video does not explain how the human demonstrations were collected, how many were used, or how the system turns them into actions on the G1. It also does not establish repeatability, safety, or performance against other training methods.
The practical next step is testing the same skills repeatedly with changed objects, positions, and unexpected failures. That would show whether Workhorse transfers a useful skill or mainly handles the filmed examples. The evidence establishes a demonstration, but not commercial availability or business deployment.