ETH Zurich’s Soft Robotics Lab Unveils Technology That Uses Robot-Hand Fingers as Legs

On September 16, 2026, ETH Zurich’s Soft Robotics Lab posted a YouTube video showing a robot hand using its fingers as legs to walk autonomously and recover from falls through reinforcement learning. In the video, the hand crawls while supporting its own weight, changes direction, recovers its posture after falling, presses a keyboard key and pushes an object. Control uses separate policies for each task, with inference performed onboard the robot.

Using a robot hand like legs could change design assumptions by giving a component intended solely for grasping the ability to move. ETH Zurich’s announcement also showed an implementation using onboard inference, indicating that the system does not require external computing resources.

The novelty of this announcement is that it treats a robot hand not as a “grasping tool” but as a body that handles movement and interaction with the environment. Rather than simply demonstrating dexterity at the fingertips, it shows the same platform supporting its own weight, changing direction, recovering from falls, pressing keys and pushing objects. This can be seen as an effort to broaden the definition of what a robot hand can do. The implementation, however, centers not on major hardware modifications but on task-specific policies learned through reinforcement learning and onboard inference, placing the research focus on control rather than mechanics. What the two have in common, however, is that research associated with ETH Zurich focuses on “soft mechanisms” and “interaction with the outside world.” This article addresses a shift in the role of the body—reinterpreting the hand as a foot—rather than fine manipulation for surgery. So while the research themes are related, the outcomes are different. The industry question is whether a robot hand designed solely for grasping can be extended into a multipurpose body capable of movement, contact and recovery. The use of onboard inference indicates the possibility of implementation without connection to an external computer. For mass production and deployment in the field, however, the next questions are how much power, weight and computing load the approach requires. It also remains unclear how well an approach based on accumulating task-specific policies will handle more complex continuous movements or environments not seen during training. We see this announcement as a sign that the research emphasis may be shifting from “dexterity at the fingertips” to “reuse of the body as a whole.” However, further papers and experiments will be needed to determine whether the functions shown in the video remain laboratory demonstrations or can be reproduced as a general-purpose control framework.

For Japan’s robotics sector, fingertip grasping mechanisms and integrated tactile and visual control can be areas of strength. But the redefinition of a hand as a foot, as in this work, requires implementation capabilities that link mechanical design and control. The demonstration highlights the importance not only of the precision of individual components but also of a control stack that brings together movement, recovery and contact.