Real robots are dumb

Actual robots, today, are a lot clumsier than popular science fiction would suggest. Performing dynamic, non-precomputed movements is still hard in the real world.

Industrial robots (producing cars or packaging things) typically perform movements by transitioning between fixed, recorded poses, or following precomputed paths.

When they take sensor information into account at all, it's usually by following hand-engineered rules like "when a y-force exceeds the threshold, stop and start moving z-wards."

Enter intelligence

MIRAI skills are trained end-to-end with data recorded on site.

This is radical in a number of ways. First of all, the computer vision self-tailors to the task, picking up the lighting conditions and geometries involved.

Secondly, trained skills are robust to a lot of variance in the input data - neural networks are good at generalization, and their performance deteriorates much more gracefully than that of hand-written programs trying to account for all eventual disturbances.

Thirdly, end-to-end training can learn movement know-how that human workers wouldn't even be able to verbalize: Guiding the robot while still feeling and seeing the environment provides a lot of subtle information to the machine learning algorithms.

Training isn’t teaching

micropsi industries controllers are not installed „on the robot“ and do not replace any low-level controllers that come with the robot. Trainable micropsi industries controllers are meant to be brought in when point-to-point-teaching can’t solve a task, and hand-programming a solution is too tedious or impossible.


MIRAI is based on the micropsi industries Behavior Development Kit (BDK), which is available for download free of charge and partly open source.

Robot interface code, camera and sensor support, and the actual learning/inference algorithms aren't part of the BDK. We will make some of these interfaces available in the future as the product matures.