computer-vision robotics ai-agents sports-ai

Learning athletic humanoid tennis skills from imperfect human motion data

zzk273.github.io

March 15, 2026

1 min read

🔥🔥🔥🔥🔥

56/100

Summary

LATENT enables the simulation of highly-dynamic tennis rallies by human athletes, showcasing versatile skills in a competitive environment. Developed by researchers from Tsinghua University, Peking University, and other institutions, it utilizes advanced AI techniques for real-time performance.

Key Takeaways

LATENT is a system designed to learn tennis skills for humanoid robots using imperfect human motion data, which consists of motion fragments rather than complete sequences.
The system successfully develops a humanoid policy that allows robots to consistently strike and return tennis balls while maintaining natural motion styles.
LATENT demonstrates effective sim-to-real transfer, enabling the Unitree G1 humanoid robot to sustain multi-shot rallies with human players.
The approach significantly reduces the difficulty of data collection by utilizing quasi-realistic data to capture primitive tennis skills.

Read original article

View story →

Coverage & discussions

Hacker News · zzk273.github.io

Learning athletic humanoid tennis skills from imperfect human motion data ↗

Community discussions

Hacker News discussion ↗

Historical count: 5 · 5 saved in sample · Historical sample

Analysis ready

Community Sentiment

Mixed

Positives

The practical use of imitation learning from human demonstrations in humanoids is ramping up, indicating significant advancements in AI training methods.
There is optimism that general-purpose physical knowledge and capabilities in humanoid robots will be demonstrated within the next year, showcasing rapid progress in AI robotics.
The potential for robotic AI instructors for children in affluent households highlights the growing applicability of AI in personalized education and training.

Concerns

The robot's movements are still described as 'robot-like' and hesitant, suggesting that current models have not yet achieved fluidity and naturalness in motion.
Estimating the robot and ball pose with onboard cameras remains a significant challenge, indicating limitations in current state estimation methods for closed-loop robotics.