Research map / Probabilistic and scientific modeling / Molecules and proteins

Prioritizing Perception-Guided Self-Supervision: A New Paradigm for Causal Modeling in End-to-End Autonomous Driving

NeurIPS 2025 · 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 Perception-Guided Self-Supervision (PGS), which uses perception outputs (lane centerlines, predicted agent motions) as positive and negative self-supervision for ego planning in a VAD-like end-to-end driver, reaching a Bench2Drive Driving Score of 78.08.

Problem

Imitation-learning end-to-end driving suffers causal confusion and poor closed-loop performance because noisy expert trajectories serve as the sole supervision target.

Approach

Replace or filter expert trajectory targets with perception-derived signals: lane-selection classification (MTPS), centerline-aligned trajectory targets (STPS), and SAT-based negative collision penalties from predicted agent boxes (NTPS), added on top of a standard VAD-style architecture.

Key results

• PGS achieves a Driving Score of 78.08 and 48.64% success rate on Bench2Drive, versus 42.35 and 15.00% for VAD-Base.
• Ablation: PGS_All reaches a mean Merging/Overtaking ability success of 54.17% versus 14.89% for PGS_Base (imitation only).
• PGS_self (no imitation loss) reaches 33.40% mean success but a much worse Avg. L2 of 2.89.

Measured effects reported in the paper

Surprising findings

• Training with self-supervision alone and no imitation loss still gives a respectable success rate of 33.40%.

Limitations

• Performance depends on accurate perception of agents and road structure.
• Failure cases include opened car doors and emergency vehicles not being distinguished semantically.
• Comfort score is low (12.37) relative to other methods.

Experimental setup

Datasets: Bench2Drive

Benchmarks: Bench2Drive

Metrics: Driving Score Success Rate Efficiency Comfortness Avg. L2

Baselines: VAD-Base VAD-Tiny UniAD-Base UniAD-Tiny SparseDrive DriveTransformer DiffAD WoTE DiFSD TCP-traj ThinkTwice DriveAdapter

Builds on: VAD BEVFormer Bench2Drive

Closest papers by meaning

• DriveHierarchy: A Benchmark for Diagnosing VLM Driving Capabilities from Open-Loop Understanding to Closed-Loop Execution (NeurIPS 2026 · similarity 0.90)
• Continuously Learning, Adapting, and Improving: A Dual-Process Approach to Autonomous Driving (NeurIPS 2024 · similarity 0.89)
• DriveAgent-R1: Advancing VLM-based Autonomous Driving with Active Perception and Hybrid Thinking (ICLR 2026 · similarity 0.89)
• $AutoDrive\text{-}P^3$: Unified Chain of Perception–Prediction–Planning Thought via Reinforcement Fine-Tuning (ICLR 2026 · similarity 0.89)
• ChainFlow-VLA: Causal Flow Planning with Vision-Language Models (NeurIPS 2026 · similarity 0.89)
• DriveDPO: Policy Learning via Safety DPO For End-to-End Autonomous Driving (NeurIPS 2025 · similarity 0.89)