RoboNaldo: Accurate, Stable and Powerful Humanoid Soccer Shooting via Motion-Guided Curriculum Reinforcement Learning

RoboNaldo, a motion-guided curriculum reinforcement learning framework, enables humanoid robots to perform accurate soccer shots with significantly improved stability and power compared to prior approaches. The system uses a three-stage training process that progresses from mimicking human motion to adapting kicks for varied ball positions and moving targets, achieving real-world performance on a Unitree G1 robot with shot errors under 1 meter from 3 meters away.

RoboNaldo represents a meaningful advancement in robotic motor control by solving a complex challenge that combines stability requirements with high-impact dynamic movements. The research bridges two previously competing RL approaches: motion-tracking methods that ensure stability but lack adaptability, and task-reward methods that struggle with exploration. By implementing a curriculum learning strategy that scaffolds from a human reference motion, the system achieves both reliable whole-body coordination and the flexibility needed for varied real-world conditions.

This work builds on years of progress in humanoid robotics and reinforcement learning, where researchers have struggled to transfer simulation success to physical systems. The three-stage progression—stable prior learning, free-kick adaptation, and moving-ball extension—represents a thoughtful engineering solution that acknowledges practical training constraints. The 48.6% reduction in shooting error and 2.96x velocity improvement over baselines demonstrates tangible technical progress.

The real-world validation on a Unitree G1 with onboard perception carries particular significance. Achieving 0.73-0.86 meter accuracy and 13.10 m/s shot velocity (approaching 60-70% of professional standards) shows the framework generates genuinely athletic performance rather than merely simulated results. This validates the transferability of learned policies from simulation to physical embodied systems operating in unstructured environments.

Looking forward, the modular interface design allowing different high-level controllers to guide the learned policy suggests potential applications beyond soccer. The framework's emphasis on curriculum learning and motion scaffolding may influence broader approaches to complex manipulation and locomotion tasks in robotics.

• →Three-stage curriculum RL framework reduces free-kick shooting error by 48.6% and increases velocity 2.96x versus prior baselines
• →Real-world Unitree G1 robot achieves 0.73-0.86m target accuracy from 3 meters in both free-kick and moving-ball scenarios
• →Hybrid approach combining motion tracking stability with task-reward flexibility solves exploration challenges in high-impulse robot movements
• →Shot velocity of 13.10 m/s reaches 59-71% of professional soccer performance levels
• →Modular design enables different high-level controllers to guide the same learned low-level policy, improving transferability