How do you use machine learning to control a drone without worrying it’ll crash or do something weird?
This group says that by separating out the deterministic behavior and using ML only on the stochastic component, they can give formal guarantees for safety and robustness.
This group says that by separating out the deterministic behavior and using ML only on the stochastic component, they can give formal guarantees for safety and robustness.
https://twitter.com/guanyashi/status/1463758745096241154
They do a few other things.
They fuse learning and control theory to derive error bounds on their (learned) controllers, ensuring the drone’s trajectory remains within a specified set.
And they show a demo that the math works via a drone that skims the ground.
They fuse learning and control theory to derive error bounds on their (learned) controllers, ensuring the drone’s trajectory remains within a specified set.
And they show a demo that the math works via a drone that skims the ground.
In addition to this continuous correction (keeping errors within a known ball in trajectory space) they do a discrete correction that forces learning to only occur subject to specific cleverly chosen invariants.
I love this kind of paper b/c the experiments show the math works.
I love this kind of paper b/c the experiments show the math works.
They use this open source programmable drone[1] for their experiments, the Bitcraze Crazyflie 2.1. You can buy it online, then script a whole fleet with the associated Crazyswarm package [2].
[1]: bitcraze.io/products/crazy…
[2]: crazyswarm.readthedocs.io/en/latest/
[1]: bitcraze.io/products/crazy…
[2]: crazyswarm.readthedocs.io/en/latest/
• • •
Missing some Tweet in this thread? You can try to
force a refresh
