A new framework allows skills demonstrated by a human to be converted into general movement strategies that different robots can perform based on their physical design. This eliminates the need to programme tasks separately for each robot, making systems more sustainable and cost-efficient.
This is Kinematic Intelligence, a framework developed by researchers affiliated with the Learning Algorithms and Systems Laboratory (LASA) at the Swiss École Polytechnique Fédérale de Lausanne (EPFL). The framework solves a key challenge in robotics, according to Aude Billard, head of LASA. Billard: "How can you transfer a learned skill to robots with different mechanical structures, while ensuring safe and predictable behaviour? This approach could significantly reduce the time and expertise needed to deploy robots in real environments."
Capturing demonstrations
EPFL developed the framework by using motion-capture technology to capture tasks demonstrated by humans, such as placing, pushing and throwing. These recordings were then converted into general movement strategies.
The researchers also developed a systematic classification of the physical limitations of different robot designs. These include joint range of motion and positions that must be avoided to maintain stability. The framework then automatically adapts these general strategies to allow different robots to perform tasks safely within their specific mechanical limits.
Successfully tested
The system has been successfully tested. During an experiment on a production line, a human demonstrated a task by pushing a wooden block from a conveyor belt to a workbench, placing it on a table and finally throwing it into a basket. Using Kinematic Intelligence, three different commercial robots then managed to perform the same set of tasks reliably.
"Each robot performed different steps of the task, and the system functioned successfully even when the distribution of steps was changed," explains Sthithpragya Gupta, PhD student at LASA and one of the authors of the study. "Each robot interprets the same skill in its own way, but always within safe and feasible limits."
Expand to more applications
The researchers aim to extend the framework to applications such as human-robot collaboration and natural language-based interaction. Kinematic Intelligence, for example, would allow a person to give simple commands to a robot at home without technical programming knowledge.
The approach is also interesting in view of the rapid development of robot platforms. With the rapid evolution of hardware, robots are being replaced relatively quickly. Thanks to Kinematic Intelligence, machine tasks can be quickly and reliably taught to new robots, without requiring a large time investment. This simplifies the transition to new robot platforms.
Durgesh Haribhau Salunkhe, scientist at LASA and one of the authors: "Our goal is to remove the need for technical expertise, while still ensuring safe and reliable operation. The user brings the idea and desired behaviour, and the robot should take care of the rest."
The results were published in Science Robotics.
