A research team has developed RoboNaldo, a groundbreaking reinforcement learning framework that enables humanoid robots to perform accurate soccer shooting comparable to professional human athletes.
The Innovation
RoboNaldo employs a three-stage motion-guided curriculum reinforcement learning approach. The framework first learns a stable whole-body kicking prior, then adapts to free-kick settings with stationary balls at random positions, and finally extends to moving-ball shooting through a locomotion-command and kick-trigger interface.
Key Results
On a real Unitree G1 humanoid robot with onboard perception, RoboNaldo achieved remarkable results:
- Free-kick error: 48.6% lower than prior work baselines
- Shoot velocity: 2.96x faster than previous methods
- Average target shooting error: 0.73m (free-kick) and 0.86m (moving-ball) from 3m distance
- Ball velocity: 13.10 m/s, representing 59-71% of reported professional open-play shot speed
Technical Approach
The system uses a single human-kick reference as a scaffold and progressively shifts optimization towards shooting performance. A high-level heuristic planner controls the interface during training, while alternative controllers can drive the same low-level policy at inference.
This research demonstrates that with proper curriculum design, humanoid robots can master complex dynamic interactions that require coordination across the entire body.
