LATENT trains humanoid tennis from human clips

Motion fragments lower the data-collection bar and avoid the bottleneck of recording idealized tennis teleop. The sim-to-real layer matters as much as the latent model, because a real Unitree G1 sustaining rallies is the real proof point. The open-source repo makes the work more than a demo for researchers who want to reuse the tracking and policy pipeline. The caveat is scope: tennis is a flashy but narrow benchmark, so broader transfer to general humanoid tasks is still unproven. The staged release in the repo suggests this is still actively being unpacked, not a finished platform.