Mini-JEPA

Compact action-conditioned JEPA world models for simple robotics tasks.

I am experimenting with Mini-JEPA, a small action-conditioned JEPA setup for robot control tasks in Gymnasium Robotics.

The project trains compact self-supervised world models from low-dimensional robot observations, then uses learned action priors and MPC-style planning to solve Fetch reaching, pushing, and pick-and-place tasks.