Driving · 2026-11-11

ForeseeDrive: Future-Token Conditioning for Anticipatory Generative Driving

Shaokai Wu, Xunchu Zhou, Pinzhi Wang, Bo Dong, Yuting Hu, Yanbiao Ji, Qiuchang Li, Xiao Cui, Qi Kong, Hongtao Lu, Yue Ding

arXiv:2611.14286PDF

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

ForeseeDrive predicts compact, query-indexed future tokens from present context and uses them to condition a generative driving planner before candidate materialization. The tokens are derived from logged future actor motion, occupancy evolution, route interactions, and TTC margins. A zero-token-gated residual adapter, candidate preservation, and a TTC-margin guard constrain drift and brittle safe-to-unsafe boundary crossings.

Why it matters

The authors argue that fixed-candidate future-aware scoring can’t recover maneuvers absent from sampled support, turning anticipation into a support-allocation problem. ForeseeDrive provides a deployable, planner-facing future representation that enters generation (not just post-hoc selection) and is bounded with exact zero-token no-op, preservation, and localized TTC-margin guarding.

Method

  • Convert logged future evolution into query-indexed teacher tokens (actor, route, occupancy, TTC-margin queries); train a deployable predictor to estimate the same tokens online from present history, map, and route.
  • Condition a pretrained generative planner before sampling using a zero-token-gated residual adapter: geometry and logit residual paths, with bias-free injection and zero-initialized scales for exact zero-token control.
  • Constrain token-induced changes with candidate preservation and a narrow TTC-margin guard that rejects candidates turning safe base trajectories into near-contact projected near-contact transitions.

Limitation

The work’s broader impact notes that future-conditioned generation can introduce new failure modes if predicted tokens are miscalibrated or if the planner overreacts to uncertain interactions, and that deployment readiness isn’t validated by the offline and benchmark evidence; it requires closed-loop validation, safety monitoring, redundancy, and human oversight.

Abstract (from arXiv)

Generative driving planners can represent diverse ego futures, but a downstream scorer can choose only among sampled candidates; it cannot recover an absent maneuver. We present ForeseeDrive, which turns future scene evolution into a deployable interface for shaping candidate generation. Rather than reconstructing dense future states or scoring a fixed menu, ForeseeDrive derives query-indexed tokens from logged actor motion, occupancy evolution, route interactions, and TTC margins, and predicts the same token space from present context at deployment. The tokens condition trajectory geometry before candidates materialize. A zero-token-gated residual path and candidate-preservation loss limit unnecessary drift, while a TTC-margin guard targets brittle safe-to-unsafe boundary crossings. Under their respective benchmark protocols, ForeseeDrive reaches 89.2 PDMS on NAVSIM navtest, 0.41\,m average L2 and 0.05\% collision rate on nuScenes validation, and a 76.50 driving score on Bench2Drive. Controlled DiffusionDrive/NAVSIM diagnostics show an exact zero-token no-op, a small aligned-over-shuffled selected-PDM advantage under preservation, mixed oracle-best quality, and localized TTC-boundary failures. Together, the benchmark results and bounded mechanism tests support future-conditioned generation as a practical interface for anticipation. Code is included in the supplementary materials.

Related papers

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-11-10 · Indexed by Pub-AI: 2026-11-11 UTC

preprint

ForeseeDrive: Future-Token Conditioning for Anticipatory Generative Driving

Shaokai Wu, Xunchu Zhou, Pinzhi Wang, Bo Dong, Yuting Hu, Yanbiao Ji, Qiuchang Li, Xiao Cui, Qi Kong, Hongtao Lu, Yue Ding

Preprint posted: 2026-11-10 · Original record ↗

arxiv
2611.14286
Explicit source versions (1)
  • v1 · 2026-11-10
  1. Preprint available

    Source date: 2026-11-10 · First detected: 2026-11-11 UTC · Historical baseline

    Source evidence (2)
Abstract · arxiv

Generative driving planners can represent diverse ego futures, but a downstream scorer can choose only among sampled candidates; it cannot recover an absent maneuver. We present ForeseeDrive, which turns future scene evolution into a deployable interface for shaping candidate generation. Rather than reconstructing dense future states or scoring a fixed menu, ForeseeDrive derives query-indexed tokens from logged actor motion, occupancy evolution, route interactions, and TTC margins, and predicts the same token space from present context at deployment. The tokens condition trajectory geometry before candidates materialize. A zero-token-gated residual path and candidate-preservation loss limit unnecessary drift, while a TTC-margin guard targets brittle safe-to-unsafe boundary crossings. Under their respective benchmark protocols, ForeseeDrive reaches 89.2 PDMS on NAVSIM navtest, 0.41\,m average L2 and 0.05\% collision rate on nuScenes validation, and a 76.50 driving score on Bench2Drive. Controlled DiffusionDrive/NAVSIM diagnostics show an exact zero-token no-op, a small aligned-over-shuffled selected-PDM advantage under preservation, mixed oracle-best quality, and localized TTC-boundary failures. Together, the benchmark results and bounded mechanism tests support future-conditioned generation as a practical interface for anticipation. Code is included in the supplementary materials.