How a Unitree G1 Can Learn Curved Soccer Shots

When a foot strikes through the center of a stationary ball, most impulse drives translation. Contact away from the center can add spin. The rotating ball then experiences aerodynamic forces that bend its path.

The robot must control foot angle and tangential velocity at the instant of impact. A few centimeters of error can change both direction and spin, which makes repeatability harder than a straight power shot.

The support leg carries the body while the kicking leg accelerates. The torso and arms counterbalance angular momentum, and the pelvis positions the foot relative to the ball. If the base shifts early, the strike point changes.

A successful controller therefore coordinates many joints rather than replaying one leg motion. It must also recover after impact because the kick transfers momentum into the ball and away from the body.

RoboNaldo uses one human kick reference to learn a stable whole-body prior, then shifts optimization toward target shooting. Its curriculum progresses from stable motion to stationary-ball free kicks and finally moving-ball shots.

This approach addresses two problems: pure imitation may not adapt to a ball in a new position, while reward-only reinforcement learning may struggle to discover a physically valid kick from random exploration.

On hardware, the robot must estimate the ball and target relative to its body. RoboNaldo reports onboard LiDAR-camera perception on a real Unitree G1. The controller then uses the estimate to decide locomotion and kick timing.

Lighting, grass, rolling motion and partial occlusion can shift the estimate. A policy may look perfect with a known simulator state and fail when the physical ball moves between sensing and contact.

RoboNaldo reports 0.73 m average target error from 3 m for real free kicks, 0.86 m for moving-ball cases and post-contact ball speed up to 13.10 m/s. Those values describe the authors’ test setup.

The source tweet’s exact curved-shot demonstration is not accompanied by those measurements or a confirmed paper link. It should not automatically inherit RoboNaldo’s numbers, team identity or training method.

A rigorous banana-kick test would publish target position, initial ball position, spin rate, trajectory, landing error, contact success and repetitions. It should compare straight and curved targets and include failed contacts.

Robot soccer is advancing because learned policies now combine dynamic balance, perception and high-impulse interaction. The credible story is the measured progression from isolated kicks to repeatable directed shots, not the claim that the robot learned every technique “by itself.”