PaperlayerView exact revision on arXiv

Source-linked AI summary

Data-Driven Risk Fields for Safer End-to-End Autonomous Driving

Yuanxin Tian, Zhiyuan Liu, Jinhao Li, Zhenhua Xu, Wenhao Yu, Jianqiang Wang

arXiv:2609.10377v1cs.ROcs.CV

TL;DR

End-to-end driving models often lack explicit ego-conditioned risk supervision, while handcrafted or heuristic risk scores do not provide a robust absolute target. DRiF learns a shared BEV representation with static mapping, dynamic risk prediction, and planning, using pairwise labels to preserve relative risk ordering; on Bench2Drive, it reports improvements in driving score, success rate, and interaction abilities.

  • Problem

    Existing end-to-end models lack explicit risk-aware learning, while absolute risk scores depend on handcrafted scales and existing supervision may not capture ego-conditioned planning risk.

  • Method

    DRiF jointly learns shared BEV features for static map segmentation, dynamic risk prediction, and trajectory planning, supervising dynamic risk with pairwise labels generated from rule-based safety priors.

  • Results

    DRiF reports improvements in driving score, success rate, and interaction abilities on closed-loop Bench2Drive evaluations.

  • Takeaways & Limitations

    Relative-risk supervision connects explicit safety structure with data-driven end-to-end planning while retaining interpretable safety structure.

  • Takeaways & Limitations

    The pairwise label generator covers only overlap, corridor, and occupancy risks, so it is not yet a fully comprehensive risk field for real-world applications.

Abstract

from arXiv · show

Safety is a fundamental requirement for autonomous driving, yet existing end-to-end driving models still lack explicit risk-aware learning capacities. Existing rule-based risk models provide interpretable safety priors, yet their absolute risk scores depend on handcrafted functions, coefficients, and thresholds. Learning-based risk representations reduce part of this manual design, but their supervision often relies on occupancy-derived labels or heuristic cost values, which may not capture ego-conditioned planning risk. In this paper, we propose DRiF, a data-driven risk-field framework for safer end-to-end autonomous driving. DRiF learns a shared BEV feature with static map segmentation, dynamic risk prediction, and vehicle planning. For dynamic risk learning, DRiF converts rule-based safety priors into pairwise risk labels, and trains the risk field to preserve relative risk ordering instead of regressing handcrafted absolute scores. Experiments on Bench2Drive show that DRiF achieves competitive overall performance, with consistent improvements in driving score, success rate, and collision-related metrics. These results establish relative risk supervision as an effective way to connect explicit safety structure with end-to-end planning. The data and code will be publicly available.

I. INTRODUCTION

Safety requires end-to-end driving models to represent how scene elements affect ego behavior, yet existing approaches often lack explicit ego-conditioned risk supervision. DRiF addresses this gap with relative-risk learning and reports competitive Bench2Drive performance across safety and driving metrics.

  • Motivation: Existing end-to-end planners learn perception-to-planning relations implicitly, without explicitly identifying dangerous BEV regions or their influence on ego planning.This limitation is especially relevant to long-tail scenarios such as cut-ins, merging, intersections, and pedestrian crossings.
  • Motivation: Rule-based safety metrics provide interpretable interaction criticality, but handcrafted functions, coefficients, and thresholds make absolute risk scores difficult to generalize.Examples include time-to-collision, time headway, post-encroachment time, and deceleration rate to avoid collision.
  • Method: DRiF decomposes risk into static map constraints and dynamic interactive risk, using dense supervision for the former and relative ranking for the latter.The dynamic field captures planning risk induced by surrounding-agent futures and potential conflicts where calibrated dense labels are unavailable.
  • Contributions: Relative-risk supervision learns driving risk through pairwise comparisons rather than handcrafted absolute scalar labels.This design preserves interpretable safety structure while reducing dependence on a fixed risk scale.
  • Contributions: DRiF learns ego-conditioned risk representations for trajectory planning and reports competitive performance across overall, safety, and multi-ability metrics on Bench2Drive.The framework explicitly models how surrounding agents influence the ego vehicle during planning.

II. RELATED WORK

End-to-end driving increasingly uses structured BEV and scene representations, while explicit ego-conditioned risk supervision remains under-explored. Prior safety approaches range from classical interaction metrics and risk fields to learning-based spatial risk representations.

  • End-to-End Autonomous Driving: BEV, online mapping, vectorized scene modeling, motion prediction, and occupancy prediction strengthen scene reasoning for end-to-end driving.These representations provide spatial, road-structure, and future-motion information for downstream planning.
  • End-to-End Autonomous Driving: Ego-conditioned risk supervision in a shared representation remains under-explored despite advances in scene modeling.This gap distinguishes explicit planning risk from general scene understanding.
  • Risk Representations for Driving: Classical safety methods quantify interaction criticality with metrics such as time-to-collision, time headway, post-encroachment time, and deceleration rate.Risk-field methods instead represent safety as spatial cost through potential fields and interaction fields.

III. PROBLEM DEFINITION

DRiF defines an ego-conditioned BEV risk field whose values describe how locations affect the current vehicle’s safety. It separates static map constraints from dynamic interactive risk because the latter lacks calibrated dense ground-truth values.

  • Ego-Conditioned BEV Risk Field: DRiF receives observations, ego state, navigation command, and navigation point, then encodes them into a discretized BEV feature grid.Continuous BEV locations x=(x,y) correspond to grid indices q=(u,v).
  • Ego-Conditioned BEV Risk Field: The risk field is ego-conditioned because each BEV location’s safety effect depends on ego state, route intention, reachability, road structure, and surrounding-agent futures.Thus, risk is not treated as an objective scene attribute like occupancy or object identity.
  • Risk Field Decomposition: Static map representation captures dense road semantics and constraints, including road boundaries, lane markings, and lane centerlines.These semantics can be supervised with absolute labels and support risk learning and planning.
  • Risk Field Decomposition: Dynamic interactive risk represents planning risk from surrounding-agent futures and potential conflicts, but its exact numerical values lack calibrated dense ground truth.A conflict region can be riskier than a free lane-center region without having an objectively defined risk gap.
  • Risk Field Decomposition: DRiF therefore defines dynamic risk through relative ordering rather than absolute values.This separates ambiguous dynamic interactions from statically supervised constraints.

C. Relative-Risk Supervision

DRiF supervises dynamic risk by comparing pairs of planning-relevant BEV locations rather than assigning absolute risk values. The resulting relative supervision provides explicit risk information without requiring a common numerical scale.

  • Pairwise Supervision: Each sampled BEV pair receives a label indicating whether the first location is riskier, safer, or comparable to the second.The labels are y_ij∈{+1,0,−1}, corresponding to riskier, comparable, and safer relations.
  • Pairwise Supervision: Pairwise labels encode the desired ordering of dynamic risk values across BEV grid locations.The supervision constrains relative risk rather than fitting a calibrated absolute score.
  • Design Rationale: Relative supervision combines explicit dynamic-risk structure with dense supervision for static constraints without forcing both onto one numerical scale.This supports ego-conditioned risk supervision for end-to-end planning.

IV. METHODOLOGY

DRiF is a risk-centric end-to-end driving framework that shares BEV features across perception, static mapping, dynamic risk prediction, and planning. Its two-stage training uses auxiliary scene supervision before jointly optimizing static map, dynamic risk, and planning branches.

  • DRiF extracts a shared BEV feature for auxiliary perception, static map segmentation, dynamic risk prediction, and trajectory planning.
  • Stage 1 pretrains the encoder with directly observable scene knowledge, while Stage 2 removes auxiliary perception and jointly trains the remaining branches.
  • The dynamic risk branch is trained with pairwise ranking labels generated from rule-based safety priors, while the static map branch preserves road-structure understanding.

B. Model Structure

DRiF conditions a shared BEV representation on observations, ego state, navigation, and route context, then predicts static structure, dynamic interaction risk, and future plans. Dynamic risk is learned through pairwise ordering derived from prioritized overlap, corridor, and occupancy risks.

  • Model Structure: The BEV encoder combines observations, ego state, navigation commands, and navigation points into a context-conditioned feature used by downstream branches.
  • Model Structure: The dynamic risk branch predicts ego-conditioned planning risk from surrounding-agent futures and potential conflicts, where dense calibrated labels are unavailable.
  • Model Structure: The planning branch predicts future trajectories while benefiting from the risk-aware BEV feature shaped by static-map and dynamic-risk supervision.
  • Relative Supervision: Risk signatures measure overlap, corridor, and occupancy risks, which are prioritized as R1 > R2 > R3 during pairwise label generation.
  • Relative Supervision: Pairs receive labels from class priority or, within the same class, detailed score differences; ties receive a neutral label.

D. Training Losses

DRiF combines planning, auxiliary perception, static-map, and dynamic-risk objectives. Its dynamic-risk objective ranks predicted risks for ordered pairs rather than regressing handcrafted absolute scores, with auxiliary perception used only during encoder pretraining.

  • Training Losses: The total objective combines planning, auxiliary, static-map, and dynamic-risk losses.
  • Static Map Loss: Static-map learning uses dense BEV semantic labels with cross-entropy and Dice losses to preserve static road-structure understanding.
  • Pairwise Ranking Loss: Pairwise ranking supervises ordered risk pairs and preserves relative risk ordering without regressing handcrafted absolute scores.
  • Stage-Wise Training: Stage 1 uses auxiliary scene supervision for encoder pretraining, whereas Stage 2 jointly optimizes planning, static-map understanding, and dynamic-risk learning.

A. Experiment Settings

DRiF is trained on multi-town CARLA data and evaluated in closed loop on all 220 Bench2Drive routes against representative end-to-end baselines. It achieves the strongest reported performance among evaluated non-expert methods, improving driving, success, efficiency, and collision-route metrics over TF++ multi-frame.

  • Training Dataset: The training data contain camera, LiDAR, ego state, route command, map information, and expert trajectories collected across multiple CARLA towns.
  • Training Dataset: Dynamic-risk labels use 300 BEV candidates per frame, combining strategy-guided and random sampling over reachable lanes with staged safety comparators.
  • Baselines: Baselines include UniAD, VAD, ThinkTwice, DriveAdapter, TF++, HiP-AD, and SimLingo, with TF++ identified as the strongest baseline.
  • Comparison Results: DRiF achieves 88.78 DS and 75.91 SR, surpassing TF++ multi-frame by 3.13 DS (+3.7%) and 6.82 SR (+9.9%).It also improves Eff. from 246.69 to 252.20 and reduces CRoute from 22.73% to 15.91%.
  • Experiment Settings: The closed-loop evaluation covers all 220 Bench2Drive routes, with reproduced baselines trained and evaluated under the same split, sensor setting, and protocol.

C. Ablation Studies

DRiF’s ablations show that static road-context supervision and dynamic relative-risk learning provide complementary benefits, while richer risk sources and pairwise supervision improve performance. The current risk-pair labels remain limited to representative, easily validated risk sources.

  • Static Map and Risk Supervision: Removing either static-map or dynamic-risk supervision reduces performance, indicating complementary contributions from road-context and interaction-aware cues.The ablations begin from the same baseline stage-1 checkpoint and use four-frame inputs.
  • Absolute Value Supervision: Relative supervision outperforms absolute-score regression by 3.37 DS, 5.46 SR, and 7.19 MA-Mean.This supports learning relative risk ordering rather than fitting manually calibrated scalar magnitudes.
  • Risk Sources: Combining overlap, corridor, and occupancy risks achieves the best performance among the evaluated risk-source variants.R1, R2, and R3 denote overlap, corridor, and occupancy risk, respectively.
  • Limitations: The current pairwise label generator covers only overlap, corridor, and occupancy risks, so it does not yet represent a fully comprehensive real-world risk field.The authors frame this simplified design as a proof-of-concept choice for CARLA and identify broader risk theories and real-vehicle data as future directions.

VI. CONCLUSION

DRiF learns shared BEV representations for map understanding, dynamic risk prediction, and trajectory planning, using pairwise labels to preserve relative risk ordering instead of handcrafted absolute values. Closed-loop Bench2Drive results report improvements in driving score, success rate, and interaction abilities, while qualitative visualizations provide interpretable planning cues.

  • VI. CONCLUSION: DRiF learns a shared BEV representation for static-map understanding, dynamic risk prediction, and trajectory planning.The framework combines these functions within a data-driven risk-field approach.
  • VI. CONCLUSION: DRiF’s qualitative visualizations highlight ego-relevant risks and provide interpretable planning cues.The visualizations complement the quantitative closed-loop evaluation with an interpretable view of the learned risk field.
  • VI. CONCLUSION: Pairwise risk labels preserve relative ordering instead of handcrafted absolute values while retaining interpretable safety structure.This design connects explicit safety structure with data-driven learning.
  • VI. CONCLUSION: Closed-loop Bench2Drive results show improvements in driving score, success rate, and interaction abilities.The benchmark table reports closed-loop results across the evaluated driving metrics and safety statistics.
Loading 2609.10377v1…