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

2026-06-09

AI Summary

This work addresses the challenge of enabling humanoid robots to perform high-precision, stable, and high-impact soccer kicks in dynamic environments by proposing a three-stage curriculum reinforcement learning framework. Starting from a single human kicking motion as prior knowledge, the approach progressively advances from shooting stationary balls from arbitrary positions to intercepting and shooting moving balls. The method innovatively integrates fixed motion references with task-specific rewards, leveraging motion tracking guidance, whole-body dynamics control, and a decoupled architecture that separates high-level heuristic planning from low-level policy execution to enable efficient training and policy reuse. In simulation, free-kick accuracy improves by 48.6% and ball velocity increases by 2.96×. On the Unitree G1 robot, real-world experiments achieve average errors of 0.73 m and 0.86 m for stationary and moving balls at 3 meters, with a ball speed of 13.10 m/s—reaching 59–71% of professional human player performance.

Abstract

Elite humanoid soccer shooting requires whole-body stability, high-impulse whole-body interactions, and accuracy to targets. Motion tracking-driven reinforcement learning (RL) provides stability in whole-body movement coordination, but a fixed reference makes it hard to adapt to varied ball positions and strike timings; in contrast, task reward-driven RL struggles to explore and discover valid kicks from scratch. We therefore introduce RoboNaldo, a three-stage motion-guided curriculum RL framework for high-impulse humanoid interaction. A single human-kick reference is used as a scaffold and progressively shifts optimization towards shooting performance. The curriculum first learns a stable whole-body kicking prior, then adapts the kick to free-kick settings where the ball is stationary at random positions, and finally extends it to moving-ball shooting through a locomotion-command and kick-trigger interface. A high-level heuristic planner controls this interface during training, while alternative high-level controllers can drive the same low-level policy at inference. In simulation, RoboNaldo demonstrates free-kick shot error 48.6% lower and shoot velocity 2.96x than prior work baselines. In real world on a Unitree G1 with onboard perception, RoboNaldo attains 0.73 m and 0.86 m average target shooting error from 3 m away in free-kick and moving-ball cases, accordingly. And the post-contact ball velocity reaches 13.10 m/s, which is 59-71% of reported professional open-play shot speed. Project page: $\href{https://opendrivelab.com/RoboNaldo}{\text{opendrivelab.com/RoboNaldo}}$.

Problem

• humanoid soccer shooting
• whole-body stability
• high-impulse interaction
• shooting accuracy
• dynamic ball handling

Innovation

• curriculum reinforcement learning
• motion-guided RL
• humanoid soccer shooting
• whole-body control
• high-impulse interaction