Research map / 3D vision / 3D generation

Aligned Delta-Triplane Transformers as Occupancy World Models

NeurIPS 2026 · 3D vision › 3D generation · approach: Scene-aware human motion synthesis

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Delta-Triplane Transformers (DTT), an occupancy world model that pretrains a triplane occupancy autoencoder and autoregressively predicts per-plane triplane deltas with multi-scale Transformers, decoding them into future occupancy and planning trajectories.

Problem

Occupancy world models for autonomous driving typically predict full future occupancy states from VAE-style BEV latents with large Transformers, which is costly, redundant, and accumulates errors over long horizons.

Approach

Represent occupancy with a compact temporal triplane latent and forecast sparse, concentrated latent changes (deltas) per plane rather than full states; the deltas also serve as sparse queries for motion planning.

Key results

• DTT-O achieves 30.85 average mIoU on 4D occupancy forecasting vs. 27.10 for DOME-O, and 74.58 average IoU vs. 36.36.
• DTT-O reaches average L2 of 1.00 m and average collision rate of 0.30% on motion planning, the best among compared methods.
• DTT runs at 26 FPS on an RTX 4090 with 903 M memory, versus 18 FPS for OccWorld.

Surprising findings

• DOME-F outperforms DTT-F at the 0s reconstruction step, which the authors attribute to FB-Occ noise and DOME's VAE-style encoder robustness.
• DTT-O underperforms OccLLaMA-O on short-term (1-2 s) planning, with the authors attributing this to small delta prediction errors affecting near-future decisions.

Limitations

• Residual decoding artifacts appear in predictions because triplane deltas preserve fine detail and add noise to previous triplanes.
• Fragmented or incomplete road boundaries in sparse LiDAR occupancy labels lead to underfitting of boundaries and planning deviations.
• Short-term planning (1-2 s) collision rate is worse than OccLLaMA-O due to small delta prediction errors.

Experimental setup

Datasets: Occ3D nuScenes

Benchmarks: Occ3D nuScenes

Metrics: mIoU IoU L2 collision rate

Baselines: OccWorld OccLLaMA RenderWorld OccLLM DOME IL NMP FF EO ST-P3 UniAD VAD

Builds on: OccWorld triplane representation DOME UNet ResNet-18

Closest papers by meaning

• COME: Adding Scene-Centric Forecasting Control to Occupancy World Model (NeurIPS 2025 · similarity 0.88)
• Driving in the Occupancy World: Vision-Centric 4D Occupancy Forecasting and Planning via World Models for Autonomous Driving (AAAI 2025 · similarity 0.87)
• OccProphet: Pushing the Efficiency Frontier of Camera-Only 4D Occupancy Forecasting with an Observer-Forecaster-Refiner Framework (ICLR 2025 · similarity 0.86)
• GaussianWorld: Gaussian World Model for Streaming 3D Occupancy Prediction (CVPR 2025 · similarity 0.86)
• Cam4DOcc: Benchmark for Camera-Only 4D Occupancy Forecasting in Autonomous Driving Applications (CVPR 2024 · similarity 0.86)
• Semi-Supervised Vision-Centric 3D Occupancy World Model for Autonomous Driving (ICLR 2025 · similarity 0.85)