recuriosity/recuriosity

This repository contains the official code for a research paper titled "Remember to be Curious: Episodic Context and Persistent Worlds for 3D Exploration," from researchers at the University of Toronto, UC Berkeley, Wayve, and affiliated institutions. The project trains an AI agent to navigate and explore 3D indoor environments using only what it can see through a simulated RGB camera, without any map or position sensor. The core idea is that the agent learns to explore by building up a 3D reconstruction of the scene it is moving through, and uses the quality of that reconstruction as a reward signal. When the agent visits new areas or views that fill in gaps in its mental model of the room, it gets a higher reward. The reconstruction system is called Gaussian Splatting, a technique that represents 3D scenes as collections of small translucent blobs rather than explicit geometry. As the agent moves, this reconstruction is updated in real time and used to score how useful the current view is. The agent itself uses a visual encoder called DINO to process camera frames, and a transformer-based policy with a sliding attention window so it can remember recent context without growing unbounded memory use. The training algorithm is PPO, a standard reinforcement learning approach that balances exploration and exploitation. All of this is built on top of the Habitat simulation platform, which provides realistic indoor 3D environments from the HM3D and Gibson scene datasets. Once trained on exploration, the same checkpoint can be fine-tuned for other tasks. The repository includes fine-tuned versions for apple picking (navigate to and collect small objects) and image-goal navigation (find a location that matches a target photograph). Pretrained checkpoints are available on HuggingFace for all three tasks. Training the main exploration model requires 8 high-end GPUs and takes about three to six days depending on configuration. The code is organized into modules for the policy architecture, environment wrappers, evaluation scripts, and training scripts, with ablation variants provided so researchers can reproduce comparisons against alternative designs such as RNN backbones or curiosity-based reward baselines.