What is real? Real: the Unitree-exact H2 body, the MuJoCo physics every frame was simulated with, both checkpoints from the actual training pipeline, and the pro rally reference (Medvedev) they are learning to track. The video is an offline render of real simulation rollouts, policy on the left, reference on the right.
Not yet: this policy does not see the ball. The finished one takes ball state as input, tracks the rally, and composes shots. What you are watching is the learning stage - the checkpoint fails about 0.18s into the hardest window, and the montage shows that instead of hiding it. Training is live; the feed panel reads the real run.
POLICY (LEFT) vs REFERENCE (RIGHT) v28 control (thrash) first, then v36 tracking attempts - one continuous rally window, real MuJoCo physics, nothing hand-animated. Early checkpoints, shown as they are.
THE POLICY THAT WILL PLAY TENNIS. THIS IS ITS FIRST YEAR.
The clip above is where the current policy actually is: untrained-stage. v28 thrashes, v36 has started tracking the reference. It still dies about 0.18 seconds into the hardest part of the rally window.
The final robot runs this same kind of policy grown up: ball state in, rally tracking, shot composition - it will see the ball, move to it, and choose strokes. What is on this page today is the architecture preview, labeled honestly. As checkpoints get real, this page gets updated.
ATH - Association of Tennis Humanoids Physics-first humanoid tennis: pro footwork, spin-aware ball flight, drills, serves, match play. Unitree H2 body, MuJoCo simulation, no shortcuts.