Source note
How to teach the same skill to different robots
Summary
EPFL's Kinematic Intelligence aims to transfer a human-demonstrated manipulation skill across robots with different kinematics without rewriting task code for each robot. The main claim is safe, predictable cross-robot execution of the same task sequence on multiple commercial robots.
Problem
- A task programmed or demonstrated on one robot often does not work on another because robots differ in joint layout, reach, motion limits, and stability constraints.
- Reprogramming each robot from scratch raises deployment cost, slows upgrades, and makes robot fleets harder to maintain.
- Skill transfer matters in manufacturing and other settings where hardware changes faster than task specifications.
Approach
- The system starts from human demonstrations of manipulation tasks such as placing, pushing, and throwing, captured with motion-tracking.
- It converts each demonstrated task into a robot-agnostic movement strategy rather than keeping a controller tied to one robot body.
- It builds a structured description of each robot's kinematic and safety limits, including joint ranges, forbidden configurations, and stability-related constraints.
- It adapts the shared movement strategy to each robot automatically so the robot can execute the skill within its own feasible motion space.
- The paper frames this as "Kinematic Intelligence" for cross-robot skill transfer and safe execution.
Results
- In an assembly-line experiment, three different commercial robots reproduced the same demonstrated sequence: pushing a wooden block off a conveyor belt, placing it on a table, and throwing it into a basket.
- The reported outcome is safe and reliable execution across all three robots, even when the allocation of task steps between robots was changed.
- The excerpt gives no benchmark table, error rate, success percentage, or runtime numbers.
- The strongest concrete claim is one-shot transfer in the sense of "Demonstrate once, execute on many," with one human-demonstrated skill used across robots with different mechanical designs.