Robust Visual Foraging Challenge @ NeurIPS 2025
Announcement - Public Beta v0.9:
Upcoming v1.0 (planned – 1 July 2025) will add:
- Standalone Python API – train with any RL framework (PyTorch, TensorFlow, JAX) without Unity ML-Agents
- Headless Linux builds optimised for SLURM / Kubernetes clusters
Welcome to the official competition page for the Mouse vs AI: Robust Visual Foraging Challenge, part of the NeurIPS 2025 Competition Track.
Biological agents like mice can robustly perform visually guided tasks in complex environments — even under fog, blur, or visual noise. Can your agent do the same?
Goal
The goal of this competition is to benchmark and improve the robustness of artificial agents in biologically inspired visual tasks. Specifically, we aim to:
- Evaluate generalization under unseen visual perturbations
- Compare artificial agent performance to real mouse behavior under identical conditions
- Provide a biologically grounded benchmark for embodied visual AI systems
- Offer an neural alignment track to assess how well models align with cortical representations
Task Overview
Train an agent to navigate a naturalistic environment and reach a visually cued target within trialtime.
- Training scene + one perturbation (fog) are provided, distance increases incrementally.
- Final evaluation uses held‑out perturbations and starts each episode at max target distance.
- The task is adapted from a real neuroscience experiment in which mice perform the same foraging task.
This setup enables comparison between biological and artificial agents under identical visual conditions.
Download & Quick Start
Choose the build for your platform:
| Platform | Instruction | Download Link |
|---|---|---|
| Windows (GUI) | Instruction | Windows v1.0 |
| macOS (GUI) | Instruction | macOS v0.9 |
| Linux (GUI) | Instruction | Linux v1.0 |
- Download the environment
- Install dependencies
- Train your first agent
- Submit your model — Package your trained models and follow the Submission Guide.
Competition Tracks
-
Track 1 — Visual Robustness
Evaluate how well your trained agent generalises to unseen visual perturbations.
Metrics: Average Success Rate (ASR), Minimum Success Rate (MSR) across all provided and held-out conditions. -
Track 2 — Neural Alignment Predict mouse visual-cortex activity from internal representations of your agent.
Metrics: Mean Pearson correlation between predicted and recorded neural responses.
Please note that the two tracks do not need to be submitted separately. All submissions will be evaluated in both tracks.
Timeline
| Date | Milestone |
|---|---|
| June 11, 2025 | Starter kit released |
| July 1, 2025 | Competition officially begins |
| Nov 1, 2025 | Final submission deadline |
| Nov 15, 2025 | Evaluation + winners announced |
| Dec 2025 | Results at NeurIPS 2025 |
Communication
For questions, updates, and discussion, please Join the Discord You can also reach us via email at robustforaging@gmail.com