Researchers led by Junlin Long have built WAM-Policy, which turns a world-action model trained for robot arms into a policy that lets a four-legged robot without an arm follow language instructions, walking on three legs while handling objects with a front leg, according to a paper on arXiv.
In a first stage, the arm model is adapted to the quadruped and distilled into a fast student that needs one video step and two action steps. In a second stage, the teacher is improved with reinforcement learning for each task, and the specialists are merged into one fast student.
On QLM-Bench, ten simulated tasks with 1,000 demonstrations generated without teleoperation, the adapted policy reached 77%, its fast student 76.7% while sampling 14 times faster, and the merged student 85%. The team also ran the fast student on a real Unitree Go2 Air.
Background
More stories about Unitree Go2
Unitree Robotics
Robot maker
Unitree Robotics is a Hangzhou company that makes four-legged robot dogs and humanoid robots, including the G1, which research teams at Apple and other labs use.



