Remember to be Curious: Episodic Context and Persistent Worlds for 3D Exploration

TL;DR AI
2 min readKey summary
Researchers introduced a curiosity-driven 3D exploration agent that keeps a continuously updated world model and episodic RGB context.
The approach improves exploration during training and generalizes better to new photorealistic environments like HM3D and Gibson.
It also transfers to downstream tasks such as object picking and image-goal navigation, where sparse rewards make exploration hard.
The work addresses a common failure mode in curiosity-based RL: losing context in persistent 3D worlds.
