A team of scientists from the École Polytechnique Fédérale de Lausanne (EPFL) in Switzerland has revealed a new framework that embeds robots with an internal understanding of their own tools and abilities, as well as their functional limitations, called ‘kinetic intelligence.’
The researchers behind the novel framework said their approach gave the robots the unprecedented, if not somewhat dystopian-sounding, ability to learn new skills from other robots regardless of their functional similarity.
While still in the experimental stage, the team said that their learning from demonstrations (LfD) framework could dramatically improve efficiency in multi-robot environments such as large retail warehouses and automotive assembly plants by reducing training times of robotic systems designed to work together.
Learning Complex Tasks Remains Challenging for Robots of Different Designs
According to a statement announcing the sci-fi-sounding framework that lets robots learn complex tasks by watching other robots, some LfD systems already exist. However, these earlier versions cannot be transferred between different robotic systems, limiting them to a single robotic platform learning from an identical ‘trained’ model.
“Most LfD methods remain tied to the robot they were trained on,” the research team explained in a published study detailing their work.
As robotic systems become more prevalent in the workforce, their complexity and capabilities are also expected to continue to increase. The research team notes that this increasing “freedom of motion” will open the door to training methods that allow differently constructed robotic platforms to learn to accomplish tasks regularly performed by robots of a different, but similarly versatile, design. However, they also note that this increasing complexity will only increase the need for simpler training methods that don’t require hours of human intervention.
“Teaching robots new skills should be as natural as showing rather than programming,” the researchers explained. “Learning from demonstration (LfD) moves toward this goal by allowing users to guide a robot or sketch a desired motion, enabling learning without writing a line of code.”
Embedding Kinematic Intelligence on Differently Constructed Platforms
To remove human trainers from the equation, the EPFL team evaluated and classified different 3-jointed robots based on joint spacing and movement constraints. After testing various autonomous learning approaches, the team designed a framework that models potential workplaces, identifies ‘feasible’ motion trajectories based on how they had previously classified the robot, and then executes a safe movement within each robot’s specific design capabilities.
“(We) embedded kinematic intelligence into the control policy, ensuring constraint-awareness and predictable behavior regardless of design variations,” they explained.
The team notes that this framework “endows robots with kinematic intelligence: an internal understanding of their own joint limits, singularities, and connectivity.”
“Instead of correcting for these constraints after learning, we embedded them directly into the control policy from the outset,” they explained. “
Tests Validate Learning From Their Demonstration Framework After a Single Observation
To validate their new framework, the team tested it across several industrial robots with varying abilities and limitations. According to the study authors, this included “diverse simulated and real robots, both redundant and nonredundant, with varied link geometries and joint configurations.”
During each test, a human ‘trainer’ would perform a specified task to demonstrate the desired result. For example, the researchers would trace out the letters of a word or manipulate objects as if they were on a simulated assembly line.
After a single human demonstration, the robotic platforms with the embedded kinetic intelligence framework showed the uncanny ability to adapt their own tools to achieve the same result, regardless of design limitations. The researchers said this included demonstrating the ability to reproduce the traced characters, and to “push, pick-and-place, and throw objects successfully.”
“The Demonstrated Skill Executes Safely and Consistently Across Robots”
When discussing the successful tests, the researchers noted that robots imbued with kinetic intelligence could observe one or more demonstrations, extract a “globally stable dynamical system,” and produce behaviors that remain valid across robots with different physical tools and limitations. They also noted that one robot could watch another perform the skill and adapt its unique configuration to achieve the same result, without accidents or damage.
“The demonstrated skill executes safely and consistently across robots without retuning, thereby achieving cross-robot skill transfer,” they explained.
When discussing possible applications of their LfD framework that lets robots learn by watching other robots, they note that the entire design was grounded in the category of three-revolute (3R) robots, “which form the building blocks of many commercial robots.”
“This classification enables a joint space policy that preserves user intent and adapts to robot-specific constraints,” they conclude.
The study “Demonstrate once, execute on many: Kinematic intelligence for cross-robot skill transfer” was published in Science Robotics.
Christopher Plain is a Science Fiction and Fantasy novelist and Head Science Writer at The Debrief. Follow and connect with him on X, learn about his books at plainfiction.com, or email him directly at christopher@thedebrief.org.