PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning A Unified Risk Map for Autonomous Driving in Partially Observable Environments

Jie Jia, Yaofeng Su, Zeyu Bao, Yun Hong, Bingzhao Gao, Zhongxue Gan, Wenchao Ding

arXiv:2605.22189v1cs.RO

TL;DR

Occlusion-aware driving must reason about hidden agents despite conservative reachability estimates and inaccurate trajectory prediction under high uncertainty. The paper combines a unified spatiotemporal risk field with diffusion-generated adversarial scenarios for risk-aware planning, outperforming an occlusion-aware baseline on Waymo scenarios.

  • Problem

    Occlusion-aware driving lacks reliable methods for reasoning about hidden agents because reachability approaches can overestimate risk and learning-based approaches struggle under high uncertainty.

  • Method

    The framework fuses traffic-flow and collision risks, generates realistic yet adversarial occluded interactions with diffusion, and learns the resulting risk map for planning.

  • Results

    The method improves minimum TTC by 0.78 times and average TTC by 1.67 times over OPBP, while reducing critical moments by 62.3%.

  • Takeaways & Limitations

    The unified risk-field framework provides a comprehensive and practical approach to risk-aware planning under partial observability.

Abstract

from arXiv · show

Occlusion-aware prediction remains a critical challenge in autonomous driving due to the inherent uncertainty of unobserved regions. Existing approaches either overestimate risk based on reachable states or struggle to predict accurate trajectories under high occlusion uncertainty. To address these limitations, we propose a unified risk map modeling and learning framework for partially observable environments. Our method integrates traffic flow risk and collision risk through spatiotemporal modeling, enabling fine-grained assessment of occlusion-induced hazards. To address the scarcity of scenarios involving occluded interactions, we introduce a diffusion-based scenario generation framework that produces realistic yet adversarial scenarios. We integrate the modeling and learning of a unified risk map into a framework that supports risk-aware planning under partial observability. Experiments on the Waymo Open Motion Dataset show that our method significantly outperforms the state-of-the-art occlusion-aware baseline, improving minimum time-to-collision by 0.78 times and average time-to-collision by 1.67 times. The proposed framework offers a comprehensive and practical solution for risk-aware planning in partially observable environments.

I. INTRODUCTION

The paper targets occlusion-aware risk assessment and planning, where reachability methods can be overly conservative and learning-based methods struggle with hidden-agent uncertainty. It proposes a unified risk field, adversarial scenario generation, and risk-aware planning, with improved TTC on Waymo occlusion scenarios.

  • Motivation: Occlusion risks beyond the visible field are needed to formulate safer autonomous-driving strategies.Human drivers often decelerate proactively, but occluded interaction events are scarce in real-world data.
  • Limitations of Existing Methods: Reachability-based methods can be overly conservative, whereas learning-based methods struggle to predict accurate hidden-agent trajectories under high uncertainty.The former lack data-driven traffic priors; the latter face uncertainty in unobserved regions.
  • Proposed Framework: The proposed framework constructs a spatiotemporal risk field that combines traffic-flow density and potential collision hotspots.This representation is designed for partially observable environments and supports fine-grained occlusion-risk quantification.
  • Proposed Framework: An automated scenario generator synthesizes realistic yet adversarial occluded interactions to address scarce safety-critical training data.The method uses diffusion-based generation to inject real-world traffic distributions into learning.
  • Results: The learned risk map is integrated into risk-aware planning under partial observability.Experiments report significant improvement over state-of-the-art occlusion-aware baselines on Waymo Open Motion Dataset scenarios.

II. RELATED WORK

Prior occlusion-aware methods use analytical reachability or data-driven prediction, but each faces important limitations in hidden regions. Scenario-generation research also underrepresents long-tail occluded interactions, motivating automated occlusion-focused generation.

  • Occlusion-Aware Prediction: Analytical approaches estimate hidden-agent states with reachability or related formal techniques but often overestimate risk and produce conservative plans.Their limitation is attributed to missing traffic priors.
  • Occlusion-Aware Prediction: Learning-based approaches predict trajectories or occupancy maps using observed-agent interactions, but blind-zone uncertainty limits precise occluded-trajectory prediction.These prediction errors can affect subsequent planning behavior.
  • Traffic Scenario Generation: Existing traffic-scenario generation methods include replay, rules, and data-driven models, but many focus primarily on real-data distributions.Earlier replayed and rule-based methods can fail to reproduce complex, large-scale behaviors.
  • Traffic Scenario Generation: Long-tail occluded interactions receive limited attention in prior scenario-generation research.This leaves a specific gap in simulating rare interactions involving hidden agents.
  • Traffic Scenario Generation: Recent adversarial generators primarily target visible-agent interactions, motivating automated generation of rare critical occluded scenarios.The cited methods leave blind-zone simulations largely unaddressed.

B. Framework Overview

The framework builds a dense spatiotemporal risk field from traffic-flow and collision risks, trains it with generated occluded interactions, and uses a lightweight predictor for risk-aware planning. Flow density and ego-trajectory overlaps are fused into a dynamic safety map.

  • Framework Overview: The framework combines occlusion risk modeling, diffusion-based data generation, lightweight risk prediction, and risk-aware driving strategy.These components form an interconnected pipeline for efficient inference under partial observability.
  • Risk Field Construction: Multimodal trajectory sets are preprocessed, stationary agents are filtered by a speed threshold, and the map is discretized into risk grids.The active-agent set focuses modeling on relevant moving hazards.
  • Risk Field Construction: Flow Risk uses predicted-trajectory spatial density to indicate where traffic is more likely to be present.Higher density corresponds to higher flow risk in the grid representation.
  • Risk Field Construction: Collision Risk identifies spatiotemporal overlaps between the ego trajectory and predicted trajectories using a distance threshold.The resulting collision events are converted into a collision-risk field.
  • Risk Field Construction: The total risk field linearly fuses flow and collision components, then applies Gaussian filtering and normalization for planning use.The fused field serves as a dynamic safety map while reducing scene-scale and traffic-density variation.

D. Occlusion Interaction Data Generation

The framework generates rare, safety-critical occluded interactions by sampling plausible agent states, producing trajectories with diffusion, and guiding them toward adversarial yet physically realistic behavior.

  • D. Occlusion Interaction Data Generation: The generator decomposes occluded-interaction synthesis into initial-state estimation and interaction-strategy simulation.It samples potential-agent states before generating and optimizing their trajectories.
  • D. Occlusion Interaction Data Generation: Potential occluded-agent states are sampled from map topology and the ego vehicle’s field of view using positions and speeds within specified ranges.Each sample provides a prior for subsequent trajectory generation.
  • D. Occlusion Interaction Data Generation: A pretrained diffusion model generates occluded interaction trajectories by predicting controls that are converted into states through a bicycle dynamics model.The model follows the DDPM framework and includes a scene encoder and denoiser.
  • D. Occlusion Interaction Data Generation: The guidance function steers generated trajectories toward rare, safety-critical corner cases while retaining naturalistic behavior from the learned diffusion distribution.The process addresses the scarcity of critical occluded interactions in the original data.
  • D. Occlusion Interaction Data Generation: The optimization balances interaction risk against lane adherence to produce adversarial trajectories subject to physical constraints.Closest-approach distance promotes conflict, while an SDF-based road term penalizes deviations from road geometry.
  • D. Occlusion Interaction Data Generation: During denoising, the objective is maximized through gradient-based updates to the noise control sequence.Updates are applied at each denoising step.

E. Occlusion Risk Prediction

The occlusion risk predictor encodes visibility and map information, fuses them with cross-attention, and decodes lane-anchored, multi-step risk predictions using a transformer architecture.

  • E. Occlusion Risk Prediction: The predictor infers lane-anchored risk scores from vectorized environment representations for efficient and localized risk inference.Lane anchors serve as queries for decoding scene features into occlusion risks.
  • E. Occlusion Risk Prediction: The input contains field-of-view rays and scene-map polylines aligned to the ego vehicle’s coordinate frame.Ray tracing encodes visible regions, while MLPs and pooling encode map attributes.
  • E. Occlusion Risk Prediction: Cross-attention fuses visibility and map features into a compact feature vector for downstream prediction.The resulting representation captures both observation geometry and scene structure.
  • E. Occlusion Risk Prediction: Lane-anchor features interact with temporal encodings and global occlusion features through attention to decode collision risk scores.An MLP then produces multi-step risk predictions.
  • E. Occlusion Risk Prediction: The model predicts risk scores along paths, smooths them into continuous 2D risk fields with Gaussian filtering, and trains with MSE against ground-truth risks.The decoder uses a multilayer Transformer for spatiotemporal risk modeling.

F. Driving Strategy

The planner generates local trajectories along global references and optimizes a composite quadratic-programming cost that incorporates occlusion risk, collision risk, smoothness, and goal-reaching terms.

  • F. Driving Strategy: Risk-aware planning generates local trajectories along global references and optimizes them through quadratic programming.The composite cost is designed for occluded environments.
  • F. Driving Strategy: The planning cost combines smoothness, reachability, predicted occlusion risk, and visible-obstacle collision terms.These terms respectively regulate acceleration, target progress, risky-region speed, and obstacle proximity.
  • F. Driving Strategy: The occlusion-risk term discourages high speeds in regions with predicted risk.This incorporates foresight about hazards beyond the currently visible field.
  • F. Driving Strategy: The planner minimizes the composite cost to generate expert-level risk-aware trajectories in occluded environments.The stated objective combines safety-related and trajectory-quality considerations.

A. Experimental Setup

Experiments use Waymo Open Motion Dataset scenes to train and evaluate risk modeling, compare planning variants and baselines, and assess computational efficiency.

  • A. Experimental Setup: The experiments use WOMD trajectories and scene maps, with each scenario lasting 9 seconds at 10 Hz and the first 8 seconds retained.The dataset provides recorded object trajectories and off-board perception labels.
  • A. Experimental Setup: The study selects 1,000 training scenes and 100 validation scenes containing potential beyond-field-of-view uncertainty and occluded ego-agent interactions.These scenes support risk-field learning and planning evaluation.
  • A. Experimental Setup: Risk fields combine sampled occluded-vehicle initial states with multimodal trajectory distributions to compute traffic-flow and potential-collision risks.The resulting grid map uses 0.5 m resolution with values scaled to [0,1].
  • A. Experimental Setup: 6.67 ms (150 FPS) is the complete model’s average inference latency on an NVIDIA RTX 4090 GPU.FOV encoding requires 1.56 ms, while the Transformer risk decoder requires 3.45 ms.
  • A. Experimental Setup: The comparison includes NOAP and O-Risk ablations alongside occlusion-prediction-based and reachability-based planning baselines.The variants isolate the effects of risk awareness and scenario generation.
  • A. Experimental Setup: The proposed method integrates a unified risk map learned from augmented data into the NOAP planning architecture.This distinguishes it from O-Risk by using the learned map during trajectory planning.

D. Evaluation Metrics

The evaluation uses time-to-collision, risk score, and critical moments to measure collision risk, overall interaction safety, and near-collision exposure.

  • TTCmin measures the average minimum time to collision across timesteps, with smaller values indicating higher collision risk.
  • TTCavg measures average collision time across timesteps and agent pairs, with larger values indicating greater overall safety.
  • Risk Score evaluates each trajectory using ground-truth risk-field labels.Lower values indicate lower probabilities of collision or close interaction with other traffic participants.
  • Critical Moments count the average frames in which the ego vehicle has less than 3 seconds before a collision.

E. Effect Analysis

The experiments assess scenario-generation quality, risk prediction, and planning performance across occluded interactions. Results show that the proposed framework localizes evolving hazards and improves safety over competing and ablated methods.

  • Ablation Analysis: TTCmin is 7.72s for the full method, compared with 4.91s for O-Risk and 3.59s for NOAP.NOAP performs worst, while training without diffusion-generated scenarios remains below the full method.
  • Planning Performance: The method improves minimum TTC by 0.78 times and average TTC by 1.67 times over OPBP, while reducing critical moments by 62.3%.These gains are attributed to a more comprehensive and stable unified risk-field representation under high occlusion uncertainty.
  • Baseline Comparison: The proposed method achieves a 76.7% higher minimum TTC and a 287% higher average TTC than SRQ-P.Its lower Risk Score may reflect SRQ-P’s conservative velocity planning, whereas the unified risk map provides more nuanced risk assessment.
  • Risk-Aware Planning: The planner proactively decelerates before an occluded vehicle becomes visible, allowing safer interaction with minimal risk exposure.
  • Scenario Generation: Generated scenarios reduce TTC by 41.2% versus the Rule-based approach while increasing OnRoad Rate by 13.6% and reducing OffRoad Distance by 56.6%.The initial-state strategy also increases interacting agents by 49.7% over the original log.
  • Risk Prediction: Risk prediction captures low initial risk, peaks at conflict points, provides early warnings, and identifies distinct high-risk zones over time.Qualitative analyses cover complex turns, side-road interactions, adjacent intersections, and multiple occluded road configurations.

V. CONCLUSIONS

The paper presents a unified framework for risk-aware planning in occluded environments, combining spatiotemporal risk modeling with adversarial diffusion-based scenario generation. Its lightweight, lane-anchored inference improves time-to-collision metrics on the Waymo dataset.

  • The framework integrates a spatiotemporal risk model with an adversarial, diffusion-based scenario generator for occluded-environment planning.A lightweight prediction network supports efficient, lane-anchored risk inference.
  • Experiments on the Waymo dataset show significant safety improvements, including substantially increased time-to-collision metrics.
  • Future work will focus on enhancing traffic priors through improved diffusion-based sampling.
Loading 2605.22189v1…