Search papers, labs, and topics across Lattice.
This paper introduces RoboNaldo, a motion-guided curriculum reinforcement learning framework designed to enhance humanoid soccer shooting accuracy and stability. By utilizing a three-stage training process that evolves from a stable kicking prior to adapting for stationary and moving balls, RoboNaldo achieves significant improvements in shooting performance. In simulations, it reduces free-kick shot error by 48.6% and increases shoot velocity by 2.96 times compared to previous baselines, demonstrating its effectiveness in both simulated and real-world environments.
RoboNaldo achieves a remarkable 48.6% reduction in free-kick shot error, setting a new benchmark for humanoid soccer performance.
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: $\href{https://opendrivelab.com/RoboNaldo}{\text{opendrivelab.com/RoboNaldo}}$.
