nyu-visionx/solaris-eval-datasets

Solaris Eval Datasets

Project Page | Paper | Github

Evaluation datasets collected via SolarisEngine to evaluate the Solaris multiplayer world model for Minecraft. Refer to Solaris repository for downloading and evaluation running instructions. You can also use it to benchmark any multi-agent action-conditioned video model.

Dataset Info

The dataset contains videos and actions for two players in Minecraft at 720p and 20 fps. You can find the action space in the training dataset page

Below are the datasets folder descriptions and how they are referred to in the paper:

• bothLookAwayEval (Memory): Both players look away and back
• oneLooksAwayEval (Grounding): One player looks away and back
• rotationEval (Movement): One player rotates its yaw
• structureEval (Building): One player builds a structure
• translationEval (Movement): One player moves left, right, forward, backward
• turnToLookEval (Consistency): Two players look in the same direction
• turnToLookOppositeEval (Consistency): Two players look in the opposite direction

Citation

@article{savva2025solaris,
  title={Solaris: Building a Multiplayer Video World Model in Minecraft},
  author={Georgy Savva and Oscar Michel and Daohan Lu and Suppakit Waiwitlikhit and Timothy Meehan and Dhairya Mishra and Srivats Poddar and Jack Lu and Saining Xie},
  journal={arXiv preprint arXiv:2602.22208},
  year={2025}
}