Projects / Self-supervised world models
Fire: artificial agency
Experiments in agents that learn on their own: streaming deep RL, unsupervised skill discovery (METRA, CSF), cognitive map learners, and TD-JEPA.
- Streaming deep RL: stream actor-critic and stream-Q with a custom eligibility-trace optimizer (ObGD) and running observation and reward normalization
- Soft actor-critic with automatic entropy tuning
- Skill discovery: METRA and CSF (mutual-information skill learning)
- Cognitive map learners: a linear version from the Nature Communications paper, a deep version, and an original hierarchical action composer
- TD-JEPA: temporal-difference JEPA, with an Oja-rule variant of the encoders