University of michigan develops humanoid robot capable of learning to traverse rugged terrain

According to foreign media reports, researchers at the University of Michigan have developed a new artificial intelligence (AI) model that can train humanoid robots for hiking. This is expected to accelerate the development of embodied intelligence, enabling them to perform tasks such as autonomous search and rescue and ecological monitoring in undeveloped areas.

Image Source: University of Michigan

Using a new artificial intelligence framework called LEGO-H, researchers trained a humanoid robot from Unitree, equipped with cameras, to plan routes in advance, avoid obstacles, maintain balance, and adjust speed and stride based on rugged terrain.

Stella Yu, a professor of computer science and engineering and the corresponding author of the research paper, stated, "Our model is the first to grant humanoid robots fully autonomous visual perception, decision-making, and motion capabilities. It can not only walk but also jump, step over, or leap according to path requirements. Previously, humanoid robots were essentially 'blind,' with each action decision relying on human operators."

Traditionally, robots learn to navigate on flat and unobstructed surfaces through pre-constructed maps and continuous human guidance, where high-level planning ("where to go") and low-level execution ("how to move") are treated as separate issues.

Kwan-Yee Lin, a computer science and engineering researcher and the lead author of the study, stated: "Integrating navigation and mobility into a single policy learning framework allows robots to autonomously develop movement strategies based on real-world situations without the need for human pre-programmed models."

In the simulation experiment, humanoid robots are deployed on unfamiliar paths and required to navigate to a specific target point. These robots are equipped only with visual input, proprioception, and simple GPS directional guidance (e.g., "the destination is 0.3 miles to the northeast") instead of turn-by-turn navigation.

Virtual robots are categorized into adult size (approximately 1.8 meters) and child size (approximately 1.2 meters). They undergo walking tests on five different types of paths, each with five difficulty levels. Evaluation metrics include path completion, safety, and performance efficiency.

Compared to robots that have perfect navigation and environmental information obtained in advance, simulated autonomous robots are on par or even superior in terms of efficiency and safety. Researchers indicate that their built-in bodily awareness helps avoid damage, and removing this capability significantly reduces the success rate of walking.

Autonomous virtual robots adjust their body posture and movement according to the terrain. For example, when encountering a narrow space, the robot learns to tilt sideways to squeeze through obstacles. They can also choose paths based on obstacles, going around high barriers and stepping over low ones; if they cannot step over, they opt for a detour.

Lin stated, "It is astonishing that virtual robots can regain balance after falling, a capability that previous humanoid robots have never demonstrated. We did not program this function; it emerged naturally during the robots' learning and interaction with the environment."

In this preliminary study, the upper body of the robot remains fixed because adding upper body movements would significantly increase the complexity of modeling. As this proof-of-concept study succeeds in leg movements, the research team is working towards achieving full-body coordinated climbing to fully utilize the robot's entire range of motion, maximizing stability, safety, and efficiency during walking.

The research team is actively working on applying these strategies to real-world humanoid robots.