Descriptions:
A research collaboration spanning Tsinghua University, Peking University, robotics company Gal Bot, and Shanghai AI Laboratory has demonstrated a humanoid robot playing full-court tennis rallies against human opponents—a task the video argues is one of the hardest benchmarks in all of robotics. The system, called LATENT (Learns Athletic humanoid Tennis skills from Imperfect humaN moTion data), runs on the Unitree G1, a 127cm, 35kg humanoid with 29 degrees of freedom.
Rather than requiring professional motion capture, LATENT was trained on just five hours of amateur tennis players recorded in a 3m×5m capture area—approximately 17 times smaller than a real court. A three-layer architecture handles motion translation from human to robot morphology, a latent action space that compresses movement into learnable essences rather than raw joint trajectories, and a high-level policy for real-time shot selection and court positioning. Sim-to-real transfer—historically a major failure point—was addressed by deliberately randomizing physics parameters during simulation, making training noisier than reality so the deployed robot adapts rather than panics.
Quantitative results include a 91% forehand success rate, 78% backhand success rate, and sustained multi-shot rallies with the robot sprinting at over 6 m/s. TheAIGRID host provides a clear, accessible breakdown of the full pipeline for viewers following advances in humanoid robotics, reinforcement learning from imperfect data, and the sim-to-real gap.
📺 Source: TheAIGRID · Published March 19, 2026
🏷️ Format: Deep Dive
