IEEE LeHome Challenge

For the IEEE LeHome Challenge, I taught a policy to fold garments in simulation. I connected physical leader arms to simulated follower arms in Isaac Sim, so my own hand movements drove the robots in the scene, and collected hundreds of teleoperation episodes that way.

success

59%

garments

12

cameras

3

parameters

51.7M

Evaluated in simulation, the policy folded at 59% success across 12 garments.

My submission is an ACT policy (Action Chunking with Transformers). A ResNet-18 backbone reads three cameras (top, left and right), alongside the 12 joint positions of the two arms, and a transformer with a variational encoder predicts the next 100 actions at a time. I trained it for 30,000 steps; the weights are on Hugging Face.