Humanoid robots are getting good at ball sports, but matching a professional athlete’s style while still hitting the ball reliably is far from solved. A new white paper presents an adaptive motion planning and tracking framework that tackles exactly this problem for tennis.
Hierarchical planner–tracker architecture
The proposed system uses a hierarchical design:
- a high-level planner generates stylistic kinematic motion (e.g., a pro‑like tennis serve or rally),
- a low-level tracker executes this motion on the humanoid with minimal interference and high stability.
Motion data comes from broadcast video and motion‑capture recordings. These sequences are:
- reconstructed into 3D,
- retargeted to the humanoid’s kinematics,
- physically corrected so the robot can track them safely under its dynamics and joint limits.
Handling the sim‑to‑real gap with speed adaptation
When a control policy trained in simulation is deployed on real hardware, tracking performance degrades due to model mismatch, delays, friction, and perception noise. To mitigate the compounding errors from autoregressive planning and noisy sensing, the framework:
- trains the tracker on randomized execution speeds,
- conditions the planner on a learned motion‑speed representation.
This coupling allows the planner to preserve motion style while flexibly adjusting speed, and the tracker to remain robust across a wide range of real‑world conditions.
Validated on real humanoid platforms
The paper reports experiments on a research humanoid and transfer to a full‑size humanoid robot, including in‑the‑wild tennis serving without any motion‑capture system. It concludes with practical, transferable lessons for engineers building the next generation of humanoid ball‑sports systems, from control architecture choices to training strategies for robust sim‑to‑real performance.
Source: IEEE Spectrum
