Learning robust perceptive locomotion for quadrupedal robots in the wild

We present a perceptive locomotion controller for quadrupedal robots that combines fast locomotion and exceptional robustness on challenging terrain.
Learn more about this work on our project website: leggedrobotics.github.io/rl-p...
Our version of paper: leggedrobotics.github.io/rl-p...
Authors: Takahiro Miki, Joonho Lee, Jemin Hwangbo, Lorenz Wellhausen, Vladlen Koltun, Marco Hutter
Published in Science Robotics: science.org/doi/10.1126/sciro...
Video by Takahiro Miki
0:00 Introduction
0:37 Hiking experiment
1:31 DARPA SubT Challenge
1:57 Monkey research
2:08 How it works
2:59 How to train
3:33 What belief state captures
4:32 Unreliable maps
5:09 Stairs
5:31 Summary

[@basujindal]

Mind blowing piece of work, especially the part where robot crosses the Soft Obstacle and Deep Snow.

[@TNERA]

Fantastic! I agree, this is a very big advance for Quadrupedal Locomotion. The ability to allow the controller to 'not trust' or otherwise discount/discredit the sensor data is a very big advance. (!) and on the lighter side... having a controller with word like: proprioception, exteroception, and having a "Belief State". That gets us all a lot closer to Star Wars -> "Uncle Owen, this droid has a bad Motivator!" :D

[@yhwhlungs]

The deep snow is very interesting. ❄️

[@blaklite4249]

You guys are smashing it and it’s only the beginning 🤪

[@DiegoAlanTorres96]

Amazing So this is smarter than Spot, right?

[@metalmonkey128]

way smarter actually

[@blakeedwards3582]

This is incredible work! Thank you for sharing this

[@cube2fox]

Now let it compete against Spot by Boston Dynamics!

[@qingqingli1406]

Awesome work ! Bring legged robot from Lab to Wild

[@fenglongsong4760]

That's really awesome!

[@stevenhong7099]

Amazing work!

[@yuelinzhurobotics]

Great job

[@Opti-Mystic]

Thanks for the update guys. He's adorable ! Now all he needs is a fake fur cover and he'd be perfect and ready to cuddle.

[@Einmachglas]

This would be perfect for expeditions in destroyed nuclear power plants like Tschernobyl or Fukushima, don't you think?

[@ianfan7739]

太牛了！

[@robertcornwell9510]

holy cow...what's the processing power required to process this. Could I do this on a raspberry PI 4 ?

[@dshcfh]

Lol you probably could it just wouldn't be very good