How robots learn to explore unfamiliar 3D spaces

Curious robots often get stuck revisiting the same places in 3D environments because they forget where they've been. This team fixed that by pairing curiosity-driven learning with two key ingredients: a constantly-updated 3D map of the explored space and an episodic memory of the agent's trajectory history. Using online 3D reconstruction and sequence models, the approach learns pure exploration on realistic indoor scenes and then transfers zero-shot to new worlds. The learned policy adapts efficiently to downstream tasks like object picking—outperforming standard RL approaches without task-specific training.