The team behind Workhorse taught a Unitree G1 humanoid to plan and execute whole‑body manipulation by watching egocentric video of people. A visual planner predicts poses for the torso, wrists and feet, while a reinforcement‑learning tracker follows those targets. On a real robot the system sorts boxes, catches a thrown box and even climbs a suitcase, and it can recover when a person pushes it. In simulation it completed box‑sorting in 77% of runs and 64% when subjected to pushes up to 40 N·s.
Why it matters
If the approach scales, everyday service robots could handle messy, contact‑rich tasks without hand‑crafted programming.
