Humanoid Robot Climbs Ladders and Balances After Single Training Session

A robot that plans where to touch

The team tested the method on Unitree's G1 humanoid robot. Instead of setting every movement ahead of time, the system first picks the spots where the hands, feet, and other body parts should touch the ground or objects. A separate controller then moves the joints so the robot reaches those spots. In real-world demos, the robot climbed ladders and stairs, stepped across stepping stones, crawled through a narrow space, jumped over obstacles, and stood up after falling.

Key numbers

• The system plans contact for 9 body parts: both hands, both elbows, both knees, both feet, and the hips.
• A lidar sensor collects about 20,000 data points, which are processed in about 30 milliseconds.
• The environment scan updates 10 times per second (10 Hz).
• In simulation tests on 8 mixed obstacle courses, the method succeeded 92.5% of the time, compared with 58.8% for a baseline method called MGMT.
• The system runs on an NVIDIA Jetson Orin NX computer.

How it works in simple terms

Most robots follow a fixed script of movements. This method uses an attention-based neural network, a type of AI that focuses on the most useful information. It looks at the terrain, the robot's own joint positions and speeds, its body direction, and the goal a user sets. It also combines reinforcement learning, where the robot learns by trial and error, with motion imitation, where it copies example movements. A lidar mapping system turns sensor readings into height information and can spot narrow surfaces like ladder rungs.

Why it matters

The robot does not need a separate training run for every type of movement or obstacle. It can adapt its contact points to new surroundings, which could help robots work in messy, uneven places. The result is still a simulation score, though. The article notes that the same success rate has not yet been shown in the real world, so this is an early step rather than a finished product.

What to watch

• Whether the success rate holds up in real-world tests, not only in simulation.
• How well the system works on other robots besides Unitree's G1.
• Whether the approach can run fast enough on small onboard computers in everyday settings.