A blind robot learned to skateboard, without a single camera

A wooden deck. Four wheels. And a completely blind metal body.

No cameras watching the edges, no laser sensors scanning the floor.

In the labs of TU Darmstadt, a researcher decided to strip the robot of vision deliberately, then teach it to ride a skateboard.

What happened instead of falling over#

The robot developed the equivalent of human proprioception.

That is the sense that tells you where your hand is with your eyes closed. You cannot see it, but you know. It relied on its internal sensors alone and on reinforcement learning, which is strict trial and error.

The result: it balanced, then performed complex manoeuvres including a pop shove-it.

The closest comparison is a child learning to ride a bike in a pitch-dark room, working only from their sense of balance and the vibrations through their body, correcting moment by moment.

In fairness: the robot still trains suspended on safety cables that stop it destroying itself. This is not a product. It is a laboratory result.

And on the other front#

In the same days, Unitree's G1 robot was moving from balance to ballistic precision.

It is not training to kick a football. It is mastering banana kicks, the curved shot.

And away from the funny crash compilations that circulate, what happens here requires real-time physics: approach angle, and orienting body and limbs precisely within a fraction of a second.

What we are actually watching#

Not machines imitating human games for show.

Engineering systems taking the physics of movement apart and rebuilding it their own way.

The difference between those two is the difference between a funny clip and a technical trajectory. The first ends in a laugh; the second ends in a factory, a warehouse or a construction site within years.

What this means for you#

If you engineer: the lesson is not about skateboarding. It is about reducing sensor dependence. A system that works from less information is cheaper, sturdier, and less likely to fail when a camera gets dirty.

If you follow robotics: the real indicator is not demos but complex dynamic movement, and this is one.

If you worry about your job: physical movement was supposed to be the last human stronghold. It changes far more slowly than software. But it changes.

In closing#

Blindness, designed in as an obstacle, proved one thing: machines no longer need to see in order to master our movements.

Common questions#

How did the robot learn to skate without cameras?

Through the equivalent of human proprioception, relying only on internal sensors and reinforcement learning, meaning strict trial and error.

Why does this matter in engineering?

It breaks an old rule that peripheral vision was required for complex dynamic movement, showing vision is one route rather than a condition.

Is this a finished product?

No. The robot still trains suspended on safety cables, and this is a laboratory result rather than a product.

What is the practical benefit of fewer sensors?

A system working from less information is cheaper, sturdier, and less likely to fail when a camera gets dirty or the light is poor.