Unitree G1 dribbles a soccer ball using only onboard LiDAR
No motion capture, no external tracking, and a safety tether that Hong says stays slack.
Chuye Hong posted a Unitree G1 keeping control of a soccer ball through slalom courses and repeated large direction changes, using whole-body motions rather than straight-line pushing.
Ball perception is fully onboard with LiDAR
The part that makes it interesting is where the sensing happens. Hong says ball perception is fully onboard with LiDAR, with no mocap and no external tracking. Plenty of legged robot ball work has leaned on a room full of cameras telling the robot where the ball is. Take that away and the robot has to find and track a small rolling object itself, while running, while its own body is pitching around.
Trained in sim, run on hardware unchanged
Hong says the policy was trained in simulation with a human-motion prior and transferred zero-shot to hardware. Zero-shot here means no extra tuning on the real robot after simulation training. The human-motion prior is what he credits for the look of it, and he describes the result as human-like motion.
He also addresses the thing anyone watching a tethered humanoid video asks. The safety tether stays slack during operation and provides no assistance, according to his post.
Chris Paxton, who is usually quick to poke at humanoid demos, replied "All onboard! Really impressive".
All onboard! Really impressive
Shown versus claimed
What is public right now is the video and Hong's description of the setup. The paper is not out. So the onboard-only claim, the zero-shot transfer and the slack tether are all things Hong says, not things anyone outside has checked. There is also no number yet on speed, success rate over repeated runs, or how the system does when the ball gets away from it.
Still, dribbling is a decent stress test for a humanoid. It couples perception of a fast-moving object with dynamic footwork, and it punishes any lag between the two. A robot that can shove a ball in a straight line tells you very little. One that can hold the ball through tight turns has to keep re-planning both where the ball is going and where its own feet land.
The G1 is Unitree's small humanoid and has become the default research platform for this kind of work, which means other labs can try to reproduce it once the paper lands. That is the useful next step. Until then, file this as a strong-looking demo from one team, with the sensing claim being the thing worth watching when the writeup appears.