Research map / Label-efficient and robust learning / Recognition and long-tail

AlignDrive: Aligned Lateral-Longitudinal Planning for End-to-End Autonomous Driving

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Dense features and foundation priors · 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 AlignDrive, a cascaded E2E driving planner that predicts a lateral drive path first and then regresses 1D longitudinal displacements along it, plus a planning-oriented augmentation that inserts agents and rescales displacement labels.

Problem

Parallel lateral-longitudinal planning in end-to-end autonomous driving fails to couple speed decisions with surrounding agent behavior along the driving path, leading to unsafe speed profiles in interactive and rare safety-critical scenarios.

Approach

Condition longitudinal planning on the predicted drive path via anchor-based regression of 1D displacements along the path, and augment training by inserting virtual agents and proportionally shrinking the displacement labels to enforce collision avoidance.

Key results

• Bench2Drive: driving score 89.07 and success rate 73.18% (single run), above HiP-AD's 86.77 and 69.09%.
• Fail2Drive generalization setting: DS 71.4, SR 62.0, HM 66.4 among RGB-only methods, above HiP-AD's HM 61.5.
• Ablation: displacement formulation plus augmentation lowers collision rate from 16.3% to 11.4% on Bench2Drive.

Surprising findings

• Augmentation yields larger gains for displacement-based planning than for waypoint-based planning, suggesting the 1D formulation better leverages augmented data.
• Re-encoding agent queries for augmentation slightly reduces driving score relative to attending to original agent queries (87.47 vs 85.82 without augmentation).

Limitations

• Inserted agents may violate road rules since insertion uses minimal rule-based constraints without road information.
• Open-loop nuScenes L2 is not best, attributed to augmentation introducing label discrepancies.
• Comfort score is lower than several baselines (16.86 vs 48.63 for SparseDrive), attributed to necessary abrupt maneuvers.

Experimental setup

Datasets: Bench2Drive nuScenes Fail2Drive

Benchmarks: Bench2Drive Fail2Drive nuScenes

Metrics: Driving Score Success Rate Driving Efficiency Comfort Collision Rate Harmonic Mean L2 collision rate (open-loop)

Baselines: PDM-Lite SimLingo VAD SparseDrive DriveTransformer Hydra-NeXt HiP-AD UniAD PlanT 2.0 PDMLite-F2D Orion TF++

Builds on: HiP-AD SparseDrive DETR-style query decoding deformable attention

Closest papers by meaning

• UniMotion: A Unified Motion Framework for Simulation, Prediction and Planning (NeurIPS 2025 · similarity 0.89)
• Drive-R1: Bridging Reasoning and Planning in VLMs for Autonomous Driving with Reinforcement Learning (AAAI 2026 · similarity 0.89)
• DriveGPT: Scaling Autoregressive Behavior Models for Driving (ICML 2025 · similarity 0.88)
• SGDrive: Scene-to-Goal Hierarchical World Cognition for Autonomous Driving (CVPR 2026 · similarity 0.88)
• $AutoDrive\text{-}P^3$: Unified Chain of Perception–Prediction–Planning Thought via Reinforcement Fine-Tuning (ICLR 2026 · similarity 0.88)
• ChainFlow-VLA: Causal Flow Planning with Vision-Language Models (NeurIPS 2026 · similarity 0.87)