A new research paper tests adaptive tennis motion on two humanoid robot platforms.
What changed
A team led by Tao Huang posted AdaPT, a motion planning system for humanoid tennis.
The paper says AdaPT learns serving and rally styles from broadcast videos.
Its design separates planning from control. The planner creates tennis-like body motions. The tracker carries out those motions on the robot.
The researchers say this split aims to keep style while preserving task performance.
The key engineering problem
The paper flags a major gap between simulation and real robots.
Tracking gets worse on physical hardware, according to the authors. Noisy perception can add more errors.
AdaPT trains the tracker across randomized execution speeds. It also uses a learned motion-speed adapter in the planner.
The authors say those steps reduce compounding errors during execution.
Hardware and deployment reality
The team reports real-world tests on a Unitree G1 humanoid.
It also deployed AdaPT on Dobot’s 1.7-meter Atom humanoid. The paper says Atom performed tennis serves in the wild without motion capture.
This is a research demonstration, not a stated commercial product launch.
The paper does not provide payload, runtime, reliability rates, or operating cost. It also does not state how often the robots completed serves.