Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Robust Planning for Autonomous Driving via Mixed Adversarial Diffusion Predictions

Albert Zhao, Stefano Soatto

TL;DR

The paper tackles robust planning for autonomous driving under adversarial, out-of-distribution agent behaviors. It introduces MAD, a framework that trains a diffusion model to predict normal agent trajectories and biases it at test time to generate adversarial predictions likely to collide with candidate ego plans, then evaluates plans using the expected cost under a mixture of normal and adversarial predictions. By sampling from the mixture distribution, the planner remains robust to adversarial actions while avoiding undue conservatism, and it extends a rule-based PDM-Closed planner to handle this mixture. Empirical results on single- and multi-agent jaywalking and red-light violation scenarios show MAD achieving the best overall performance and notable improvements over risk-sensitive and safety-constraint baselines, without relying on offline adversarial scenario data. The approach offers a practical, generalizable way to enhance autonomous driving safety in online planning contexts.

Abstract

We describe a robust planning method for autonomous driving that mixes normal and adversarial agent predictions output by a diffusion model trained for motion prediction. We first train a diffusion model to learn an unbiased distribution of normal agent behaviors. We then generate a distribution of adversarial predictions by biasing the diffusion model at test time to generate predictions that are likely to collide with a candidate plan. We score plans using expected cost with respect to a mixture distribution of normal and adversarial predictions, leading to a planner that is robust against adversarial behaviors but not overly conservative when agents behave normally. Unlike current approaches, we do not use risk measures that over-weight adversarial behaviors while placing little to no weight on low-cost normal behaviors or use hard safety constraints that may not be appropriate for all driving scenarios. We show the effectiveness of our method on single-agent and multi-agent jaywalking scenarios as well as a red light violation scenario.

Robust Planning for Autonomous Driving via Mixed Adversarial Diffusion Predictions

TL;DR

The paper tackles robust planning for autonomous driving under adversarial, out-of-distribution agent behaviors. It introduces MAD, a framework that trains a diffusion model to predict normal agent trajectories and biases it at test time to generate adversarial predictions likely to collide with candidate ego plans, then evaluates plans using the expected cost under a mixture of normal and adversarial predictions. By sampling from the mixture distribution, the planner remains robust to adversarial actions while avoiding undue conservatism, and it extends a rule-based PDM-Closed planner to handle this mixture. Empirical results on single- and multi-agent jaywalking and red-light violation scenarios show MAD achieving the best overall performance and notable improvements over risk-sensitive and safety-constraint baselines, without relying on offline adversarial scenario data. The approach offers a practical, generalizable way to enhance autonomous driving safety in online planning contexts.

Abstract

We describe a robust planning method for autonomous driving that mixes normal and adversarial agent predictions output by a diffusion model trained for motion prediction. We first train a diffusion model to learn an unbiased distribution of normal agent behaviors. We then generate a distribution of adversarial predictions by biasing the diffusion model at test time to generate predictions that are likely to collide with a candidate plan. We score plans using expected cost with respect to a mixture distribution of normal and adversarial predictions, leading to a planner that is robust against adversarial behaviors but not overly conservative when agents behave normally. Unlike current approaches, we do not use risk measures that over-weight adversarial behaviors while placing little to no weight on low-cost normal behaviors or use hard safety constraints that may not be appropriate for all driving scenarios. We show the effectiveness of our method on single-agent and multi-agent jaywalking scenarios as well as a red light violation scenario.
Paper Structure (19 sections, 7 equations, 1 figure, 3 tables)

This paper contains 19 sections, 7 equations, 1 figure, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methods
  4. Problem Definition
  5. Preliminaries: Diffusion Models for Motion Prediction
  6. Adversarially Biasing Motion Prediction
  7. Mixture of Adversarial and Normal Predictions for Planning
  8. Extending the PDM-Closed Planner
  9. Implementation Details
  10. Results
  11. Evaluation Metric
  12. Evaluation Scenarios
  13. Baselines
  14. Overall Results
  15. Single-Agent Jaywalking Benchmark
  16. ...and 4 more sections

Figures (1)

  • Figure 1: Overview of Our Method (best viewed in color at 3x zoom). We first train a diffusion motion prediction model to take in scene context and output normal agent behaviors. We then bias the diffusion motion predictor to predict adversarial agent behaviors for the candidate plan. Finally, the planner computes the expected cost for the candidate plan, using both the normal and adversarial predictions. By taking into account both types of behaviors, our method causes the planner to be robust to adversarial behaviors but not overly conservative.