This humanoid robot climbs ladders and vaults obstacles with one AI system

This humanoid robot climbs ladders, crosses obstacles, and recovers from falls using a single learned policy.

It's Unitree's G1, running a single control system from ETH Zurich.

The trick? It predicts where its hands, elbows, knees, feet and pelvis should touch next.

A second network moves the joints to make those contacts, without losing balance.

Its head-mounted LiDAR maps the terrain, processing 20,000 points in 30 milliseconds.

It learned from just 13 minutes of human motion, then trained on random virtual terrain.

In simulation, it cleared eight mixed obstacle courses 92.5 percent of the time. The comparison method managed 58.8.

The catch? Those are simulation numbers, and it can't hang or grab things overhead yet.