ETH Zurich Team Reportedly Demonstrates Unitree G1 Climbing and Crawling

humanoid.guide reported on October 25, 2026, that an ETH Zurich research team had Unitree G1 walk, climb and crawl. The article cites an X post by RoboHub and says the robot used a single policy without manually specified transitions between motions.

If humanoid.guide's report is accurate, the demonstration points to an effort to handle multiple forms of locomotion with a single policy, rather than relying on manually specified transitions between motions, on Unitree G1. The reproducibility of recovery motions reportedly absent from the human demonstrations, and safety in the event of failures on the physical robot, are issues in assessing the approach's scope.

The technical focus of the report is the division of labor between a policy that learns how to make contact, rather than treating walking, climbing and crawling as separate sequences of motions, and reinforcement learning in simulation to adapt to terrain. Human demonstrations are said to be used to learn contact patterns, while at runtime the robot determines the placement of its hands, feet and knees based on sensor data and goals. This describes a design different from simply replaying motion demonstrations, but the article alone does not establish generalization to real-world conditions or safety. In a separate ETH Zurich study covered by this newspaper, ETH research team reports biohybrid robot using mouse-derived muscle. That study focused on an actuator combining muscle tissue and an artificial skeleton, making its subject and technical layer different from the locomotion policy for Unitree G1 reported here. A shared connection to ETH Zurich is not enough to infer continuity between the research or technology transfer. Meanwhile, Researchers create dataset for robot learning that records human activity addresses a separate issue: the collection and composition of human activity data for training. There is no information that the 13.2 minutes of motion data mentioned in this report are related to that dataset. If the description of the demonstration is confirmed as a reproducible result, robot-control evaluation would extend beyond the success of individual motions to whether a policy remains reliable across different contact conditions. However, the reported “recovery” motions not included in the demonstrations do not prove general ability to handle unfamiliar situations. The article does not specify success rates, behavior during failures, terrain or ladder conditions, number of trials, or the scope of evaluation on the physical robot. Key points to verify next include the conditions for transferring from simulation to the physical robot, evaluation methods for each motion, and ways to stop or recover if contact is misjudged. See the original source for details.