Research map / Video / Video generation
Drift-Resistant Navigation World Model with Anchored Epipolar Guidance
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 DR-NWM, which predicts sparse future anchors and fills intermediate frames via a diffusion transformer conditioned on past and future anchors with bidirectional epipolar attention masks, reducing perceptual and geometric drift in navigation world models.
Problem
Rollout-based navigation world models suffer from perceptual drift (error accumulation under autoregressive generation) and geometric drift (predictions misaligned with the agent's motion), with prior fixes relying on costly 3D supervision or expensive rollout training.
Approach
Replace sequential rollout with anchor-guided rollout: predict sparse future anchors first, then generate chunk frames conditioned on past context and the future anchor, and use intersections of bidirectional epipolar lines from feature matches to form attention masks, injecting geometry without explicit 3D supervision.
Key results
- Across four datasets and horizons 1s-16s, DR-NWM reduces FID versus NWM, e.g., 76.898 vs 140.292-scale comparisons; at 16s on RECON FID is 76.898 vs NWM 112.679 (31.8% gain).
- MEt3R on RECON at 4s skip: 0.416 for Ours vs 0.463 for NWM.
- Planning with CEM on RECON: ATE 2.78 vs 3.18 for NWM (CEM+NWM), and NMD+Ours ATE 2.70 vs 2.97.
Surprising findings
- Anchor-based inference alone improves visual quality but does not directly improve geometric error; the ablation shows geometry gains depend on the AC-DiT design.
- Epipolar masking without future conditioning yields worse Sampson error than the full model, degrading from intersection-based to line-based constraint.
Limitations
- Relies on sparse feature matching (LoFTR) and RANSAC fundamental matrix estimation, which can fail and disable the mask.
- Anchors are predicted at fixed intervals; the evaluation is restricted to four navigation datasets at low resolution (160x120 for HuRoN, 224x224 generation).
Experimental setup
Datasets: RECON HuRoN SCAND TartanDrive Ego4D
Benchmarks: RECON HuRoN SCAND TartanDrive
Metrics: LPIPS DreamSim FID epipolar distance Sampson error MEt3R ATE FDE RPE
Baselines: NWM EgoWM CEM NoMaD
Builds on: Navigation World Models (NWM) Diffusion Transformer LoFTR DUSt3R
Closest papers by meaning
- Navigation World Models (CVPR 2025 · similarity 0.87)
- Predictive but Not Plannable: RC-aux for Latent World Models (NeurIPS 2026 · similarity 0.86)
- WorldVLN: Autoregressive World Action Model for Aerial Vision-Language Navigation (NeurIPS 2026 · similarity 0.86)
- MindJourney: Test-Time Scaling with World Models for Spatial Reasoning (NeurIPS 2025 · similarity 0.86)
- LoopNav: Benchmarking Spatial Consistency in World Models (NeurIPS 2026 · similarity 0.85)
- Sparse Planning in Visual World Models via Cost Gradients (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.