Researchers from Galaxy General and Tsinghua University have developed a novel approach called LATENT to enable humanoid robots to learn tennis skills using imperfect human motion data. Instead of relying on perfect, complete datasets, the system learns from fragmented and imprecise human movements, focusing on fundamental body mechanics. This approach trains a latent action space, akin to an "action dictionary," allowing robots to select and adapt pre-learned movement patterns rather than calculating every joint movement from scratch. The system incorporates a Latent Action Barrier (LAB) to ensure smooth and stable movements, and employs domain randomization in simulation to improve real-world performance by introducing uncertainties in robot and ball dynamics. AI
IMPACT Enables robots to learn complex motor skills from real-world, imperfect demonstrations, potentially accelerating embodied AI development.
RANK_REASON The item describes a research paper presented at a conference detailing a new method for robots to learn skills from imperfect data. [lever_c_demoted from research: ic=1 ai=1.0]
- G1
- Galaxy General
- Galbot
- IROS 2026
- Mario Tennis
- Mario Tennis Aces
- Peking University
- Shanghai Artificial Intelligence Laboratory
- Shanghai Institute of Artificial Intelligence
- Top Spin
- Tsinghua University
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
