Research map / Probabilistic and scientific modeling / Molecules and proteins

DriveFuture: Future-Aware Latent World Models for Autonomous Driving

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Multimodal trajectory prediction · 0 citations (OpenAlex)

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 DriveFuture, a latent world model that conditions a diffusion trajectory planner on predicted future BEV latents (grounded to GT future during training), reporting SOTA on NAVSIM v1 and v2.

Problem

Existing latent world models for autonomous driving treat future latent states as prediction targets or auxiliary signals rather than as conditions for current trajectory planning, entangling current and future features.

Approach

Predict future latent from current latent and trajectory token, refine it via cross-attention against the (stop-gradient) GT future BEV latent during training, anneal from GT-grounded to self-predicted future condition (LatentAlign), and feed it into every denoising step of a diffusion planner, with progressive foresight guidance (kinematic and Tweedie surrogates) at inference.

Key results

• 55.5 EPDMS on NAVSIM-v2 navhard, ranked 1st on the leaderboard (vs. DrivoR 54.6).
• 89.9 corrected EPDMS on NAVSIM-v2 navtest (vs. Latent-WAM 89.3).
• 90.7 PDMS on NAVSIM-v1 navtest (vs. GoalFlow 90.3).

Surprising findings

• Direct MSE supervision of the future latent (32.1 EPDMS) underperforms implicit future conditioning (34.6 EPDMS) in the ablation.

Limitations

• Depends on predicted future latents at inference, so inaccurate forecasts can hurt planning in uncertain scenes.
• Focuses on near-future conditioning only (longer horizons not explored).

Experimental setup

Datasets: NAVSIM NavTrain nuPlan OpenScene

Benchmarks: NAVSIM-v1 navtest NAVSIM-v2 navtest NAVSIM-v2 navhard

Metrics: PDMS EPDMS

Baselines: World4Drive DrivoR DiffVLA DiffusionDriveV2 Latent-WAM DriveWorld-VLA GoalFlow DriveLaW DiffusionDrive TransFuser

Builds on: TransFuser DiffusionDrive GTRS NAVSIM

Closest papers by meaning

• Future-Aware End-to-End Driving: Bidirectional Modeling of Trajectory Planning and Scene Evolution (NeurIPS 2025 · similarity 0.90)
• Driving in the Occupancy World: Vision-Centric 4D Occupancy Forecasting and Planning via World Models for Autonomous Driving (AAAI 2025 · similarity 0.88)
• DriveGPT: Scaling Autoregressive Behavior Models for Driving (ICML 2025 · similarity 0.87)
• OneVL: One-Step Latent Reasoning and Planning with Vision-Language Explanation (NeurIPS 2026 · similarity 0.87)
• LaST-VLA: Thinking in Latent Spatio-Temporal Space for Vision-Language-Action in Autonomous Driving (NeurIPS 2026 · similarity 0.87)
• Hybrid-Driving: An Autonomous Driving Decision Framework Integrating Large Language Models, Knowledge Graphs and Driving Rules (AAAI 2025 · similarity 0.86)