Breaking

GaussianWorld Code Released, Lifting 3D Occupancy mIoU by About 2.3 Points

July 3, 2026 at 02:56 EDT

Robotics
Research & Papers
Open Source

Code has been released for GaussianWorld, a method that reframes 3D occupancy prediction for autonomous driving as a "world model" task, reportedly raising mIoU by about 2.3 points over a single-frame approach with no added computation.

December 2024 · CVPR 2025 · Tsinghua University

GaussianWorld: A World Model for Streaming 3D Occupancy

By explicitly modeling how a driving scene evolves over time with 3D Gaussians, the model lifts prediction accuracy by more than two points on nuScenes — with no extra computational cost.

+2.28

mIoU points gained on nuScenes

0×

extra compute cost added

motion factors modeled explicitly

Accuracy on nuScenes (SurroundOcc annotations)

mIoU — higher is better

19.85

Single frame
25600 Gaussians

GaussianFormer

22.13

Streaming
world model

GaussianWorld

How scene evolution is decomposed

Ego alignment

Compensate for the vehicle's own motion.

→

Dynamic motion

Track local movement of moving objects.

→

Completion

Fill in newly observed regions.

Static objects stay fixed while only dynamic ones move — so the model streams continuously instead of re-fusing stacks of past frames.

What draws attention

Higher accuracy at no added compute, plus the practicality of streaming prediction. Code, configs and trained weights are openly released.

What needs verifying

Limits in complex dynamic scenes and the scalability of the 3D Gaussian representation remain open, with real-world usage reports still limited.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.