Teaching robots new skills once required specialist programming knowledge, but that barrier may soon become obsolete. A new generation of robots is emerging—ones that can learn by observing and imitating human actions. Engineers are now developing robotic systems capable of “learning from demonstration,” which allows users to guide robots through tasks in a more intuitive manner. These training methods include remote control using a joystick, physically guiding the robot’s movements, or performing the task while the robot watches and mimics the behaviour. Each of these approaches supports a more accessible and adaptable path to robot training.
Typically, robots are designed to be trained through just one of these demonstration styles. However, engineers at MIT have developed a breakthrough interface that unifies all three methods into a single training system. This new interface, referred to as a “versatile demonstration interface” (VDI), is a compact, sensor-equipped tool that attaches to the arm of a standard collaborative robot. With this device, users can choose to train the robot through remote operation, physical manipulation, or live demonstration, depending on the task at hand or their personal preference. The goal is to make robot training more natural and less reliant on technical skills.
To assess its effectiveness, the MIT team tested the VDI on a commercial robotic arm at an innovation centre where manufacturing professionals explore technologies designed to improve industrial processes. Volunteers with factory experience were asked to train the robot to perform two common tasks: press-fitting, where pegs are inserted into holes, and molding, which involves rolling a dough-like material around a rod. Each volunteer used all three training approaches in sequence, giving researchers insight into usability and preference.
Participants generally favoured the natural teaching method, where they performed the task themselves while the robot observed. However, each method had its strengths. Remote operation was noted as particularly useful in scenarios involving hazardous materials, while kinesthetic training was found to be effective when handling tasks that required physically demanding or large objects. The study highlighted how different training styles could complement each other in various industrial contexts, making the case for a versatile interface that accommodates all three.
The VDI is equipped with a camera and sensors that record motion and applied force during training. When attached to the robot, it enables teleoperation or guided manipulation. When detached, it can be used directly by a person to perform the task, recording all necessary data for the robot to learn by imitation. This flexibility not only simplifies robot programming but also expands the pool of individuals who can teach a robot, opening the door to workers without coding expertise to contribute to robotic systems on the fly.
MIT’s researchers see broader potential beyond factory floors. The interface could be used in homes, hospitals, and caregiving settings, allowing robots to assist with a range of tasks by learning from the people they support. As postdoctoral researcher Mike Hagenow notes, “We are trying to create knowledgeable and skilled teammates that can effectively work with humans to get complex work done.” With continued development and user feedback, the team hopes that this adaptable approach to robot training will promote the broader adoption of collaborative robots across diverse industries.
More information: Michael Hagenow et al, Versatile Demonstration Interface: Toward More Flexible Robot Demonstration Collection. DOI: 10.48550/arXiv.2410.19141