Human Moves, Robotic Grooves: AI Mimics Human Motion in Stunning Detail

Learn how to use Deep Reinforcement Learning (a method of Machine Learning which involves rewarding the agent for doing something correctly, and punishing it for doing anything incorrectly) and Adversarial Motion Priors (Motion Imitation from Observations) for training robots lifelike movement in a simulation, with Boston Dynamics Atlas. Similar techniques are applicable to other humanoid robots, like Agility’s DIGIT used in Amazon warehouse or Tesla's Optimus - or even non-humanoid robots, such as #disney Imagineering latest adorable biped robot, currently being tested at Galaxy's Edge.
While I do not have exact information on how Boston Dynamics Atlas works, it is a good educated guess :)
Github:
github.com/AIWintermuteAI/Isa...
Will likely merge to main branch with time.
Motion capture files:
mocap.cs.cmu.edu/
sites.google.com/a/cgspeed.co...
Robot descriptions:
github.com/robot-descriptions...

[@Hardwareai]

Hey, guys! This is a remastered version of my video about AMP and Atlas - I put a lot of effort into making the flow of the video smoother, removed many filler words and added more visuals. Hope both you and our algorithmic overlords from KZbin like it

[@exploring-electronic]

Thanks! Good job on improving the pace of video, much easier to listen

[@exploring-electronic]

When is the next video btw?

[@Hardwareai]

I'll publish it either this weekend or next week!

[@Hardwareai]

Uploaded now!

[@therobotstudio]

Fantastic stuff!

[@Hardwareai]

Glad you enjoyed it!

[@BogdanMnikov]

how does AMP compare to other motion imitation techniques?

[@Hardwareai]

Glad you asked! There was a follow-up research, called Adversarial Skill Embeddings, which is even more promising, since with it you can incorporate a large array of different skills into one policy and seamless transitions between them. Implementing it is on my radar.