The robots were able to reproduce human skills simply by observing us.

New Way to Teach Robots from EPFL

Scientists from Lausanne (EPFL) found a solution to the problem: how to get different robots to perform the same action without rewriting the program for each type of machine. Their approach is called *kinematic intelligence*—a new form of learning-from-demonstration.

What Changed
- Learning by Observation

Instead of manually coding for each model, the robot simply watches a human perform the task.

- Adaptive Strategy

Observed movements are converted into an action plan that takes joint limits and other physical characteristics of the specific robot into account. Thus each robot “translates” the same action to fit itself while staying within safe bounds.

- Globally Stable Dynamic System

With it, even a single observed action can be reproduced correctly on different machines.

How It Works
1. Data Collection – researchers recorded people moving objects by placing, pushing, and throwing them.

2. Constraint Classification – they identified the balance limits and joint motion ranges for each robot.

3. Information Integration – the human movement data were linked to a constraint model, creating an observation and adaptive learning system.

4. Experiments – three commercial robots learned to:

- Move wooden blocks from a conveyor onto a work platform,

- Place them on tables,

- Throw them into containers.

What’s Next
The next step is training robots to perform tasks based on *verbal* instructions. The user will describe the desired behavior, and the robot will figure out how to implement it while considering its own capabilities and constraints.

Result:

Kinematic intelligence enables rapid skill transfer from humans to various robots without manual code tuning, saving time and money in research labs and industry.