Gray
quadruped

I worked on Gray for a long time and got further than I expected. I rebuilt the whole software stack and taught it to walk in simulation, and then I decided to stop building this version.

The servos are what hold it back. A hobby servo only takes a position. It can't sense or control force, and I don't think it can survive the impacts a trot puts through it. Every policy I train is limited by that, and I don't see a reason to finish a robot dog with joints that weak.

Things I learned from it:

• Writing reward functions and tuning them until the walk came out right
• Setting up a reinforcement learning pipeline and getting it to train fast on limited hardware, with 64 robots at once
• Building the simulation from accurate CAD, because the policy only learns the robot it's given
• Checking each policy against a list of pass or fail criteria
• Training separate policies for flat ground and for slopes

Next I want to build Gray 2.0. It'll use the same learning setup and reward functions on a new quadruped with quasi-direct-drive actuators. Those are brushless motors with a low-ratio planetary gearbox, the same kind of joint the MIT Mini Cheetah uses. They can control torque directly, they can be back-driven, and they handle impacts much better than servos.

I split this project into parts. Each part below has its own page, with its own sections and pictures.

Soon you'll be able to control the robot yourself, right here on this page.

The frame below is where it will go. It's a work in progress.