We introduce a system for fine-grained robotic manipulation! 🤖 What’s new? * We can control cheap robots to do surprisingly dexterous tasks * New technique that allows robots to learn fine motor skills A short thread 🧵

First, the hardware: We use simple puppeteering. Just copy the joint angles from the leader to the follower robot. No tactile or force feedback. The manufacturer (@trossenrobotics) didn’t know these tasks were possible.

Can the robot do all this on its own? We train the robot to predict actions in chunks, rather than one at a time. Recipe: action chunking + transformers + only 50 demonstrations The robot *autonomously* completes fine manipulation skills. ⬇️

But wait, there’s more! Website & Code: Paper: You can buy the full kit directly from Trossen: Led by the amazing @tonyzzhao, with @svlevine, @Vikashplus

Great to see it being finally released. Looks very cool. Congrats @chelseabfinn and @tonyzzhao !!

This looks super cool, congrats!! To what extent do the policies generalize (e.g. object variations, positions)?

Thanks @danijarh! The demos include ~15cm of object position variation, and the policies generalize to that degree. Need more & wider demo data for more generalization. We also see some generalization to distractors & disturbances.

This is mind blowing!! 🤩! Congrats!!

@LukeW I'm having the fastest hosting and utility design. Who needs can contact:

@futuristufuk

Impressive!