I've been experimenting with attention mechanisms to design locomotion controllers that adapt to different shapes, this is the same controller running on two different bodies paper coming soon 👀

my usual approach is to train MLPs that map proprioceptive inputs to muscle control signals but one limitation of this MLP approach is that it expects a fixed number of inputs and outputs, so it's not ideal for universal controllers

I implemented an attention mechanism to support any # of inputs/outputs, wrote some Python utils for originally built for the web (WASM), but I've been working on native builds for PyTorch integration to make it easier to try different architectures

the code that runs the attention-based controller is implemented as a GitHub workflow, no installation needed, can be triggered from the Actions UI if you fork the repo set how many steps to run and the video will be saved as a workflow artifact

the current setup is minimal, just two shapes (biped and quadruped), but it'd be interesting to bring in ideas from morphogenesis to explore how behavior could adapt as the body grows and changes shape... still a lot to figure out!

The Art of Flight is a homage to 80s/90s arcade action shmups with a fresh twist on the genre. Pilot multiple ships at the same time to take on oncoming waves of enemies in this fast paced space shooter. Wishlist on Steam today!

Very nice! I take it these are point simulations and the edges are springs with the red color indicating an internal force? What and how many rewards functions did you need for the gaits you show here?

reward = horizontal displacement, nothing too fancy; there is a rest length multiplier (<= 1) that basically reshapes the potential energy function defining new energy minima, red segment means multiplier closer to 0, the pink plot here is the potential energy of the muscle

@_AndrewZhao So awesome. Are you using DRL to learn the locomotion control?

@_AndrewZhao yes, I already had these controllers from old experiments where I used DRL, here I just reused them to generate training data to merge both brains into a single controller that can adapt to both bodies

this is so cool

Good work! I think figuring out how to make shapes recover from out of balance poses is a good direction to make it more robust thus more useful💪