Embedded Files
Motivation: Objective Conflict in Simulation
Motivation: Objective Conflict in Simulation
How can we successfully balance the realism and controllability in generating this safety-critical scenario?
The illustrated scenario involves Car 13 executing an unprotected left turn, prompting Car 7 to change lanes and interfere with Car 5.
Baseline methods like STRIVE, TrafficSim, and CTG fail by either generating unrealistic or non-safety-critical scenarios. CCDiff manages to generate both safety-critical and realistic scenarios.
Method Overview: Causal Composition Diffusion (CCDiff)
Method Overview: Causal Composition Diffusion (CCDiff)
Experiment Results
Experiment Results
Multi-agent Generation
Multi-agent Generation
Long-horizon Generation
Long-horizon Generation
Ablation Studies
Ablation Studies
We ablate different modules of CCDiff by changing the structured-aware encoder, causality-aware masked guidance, and different ranking strategies. The results show that ranking has the highest impact to the final performance.
Qualitative Results
Qualitative Results
We evaluate the safety-criticality of synthetic scenarios, where a higher level of accident proneness indicates greater suitability for challenging simulation in autonomous driving testing. The results show that CCDiff successfully generates realistic, safety-critical scenarios.
Ground Truth
CCDiff
CTG
BITS
Ground Truth
CCDiff
CTG
BITS
Page updated
Google Sites
Report abuse