Researchers at ETH Zurich in Switzerland have developed an artificial intelligence system that allows a robot to use virtually its entire body to overcome obstacles. In experiments, a Unitree G1 humanoid robot successfully climbed ladders, traversed narrow platforms, navigated past barriers, and regained its footing after falling. But what makes this technology so different?

Typically, teaching a robot to perform complex movements requires developing specific controls for different situations—one strategy for walking, another for climbing, and yet another for getting back up. The new system, called Attention-based Contact Planning (ACP), takes a different approach: instead of pre-determining every movement of the arms and legs, it identifies where the robot should place different parts of its body. Think of a person attempting to climb a wall; before moving their entire body, they must find secure points of contact for their hands and feet—and, depending on the difficulty, they might also use their knees or elbows.

The ACP applies a similar principle: a neural network analyzes the terrain, the robot’s position, and the target it needs to reach. Based on this information, the system plans contact points for nine body parts—including hands, feet, knees, elbows, and the pelvic region—after which another controller coordinates the joints to execute the necessary movements. To develop this capability, the researchers used just over 13 minutes of human movement data as an initial reference, combined with reinforcement learning.

In this process, the robot learns to improve its decision-making through trial and error in simulated environments. The results were significant: across eight virtual courses featuring various obstacles, the system achieved an average success rate of 92.5%, compared to approximately 58.8% for the benchmark method. While limitations remain—particularly regarding three-dimensional perception and tasks requiring more complex movements, such as hanging from structures—this research nonetheless represents an important step forward.

If humanoid robots are to operate on construction sites, in industrial facilities, or in disaster-stricken areas, walking on flat surfaces will not be enough; they must learn how to use their own bodies to navigate unexpected situations. It is precisely this capacity for adaptation that the researchers are working to develop.