Autonomous Vehicle Planning in Occluded Merges with Stochastic Safety Constraints
Year: 2024
Type: article
Abstract: We develop a robust reinforcement learning-based planning approach that handles uncertain observations and occlusions in driving scenes while meeting safety-critical constraints with a specified confidence level. We use noise models to simulate observation uncertainty and ray-tracing for occlusion modeling. The resulting uncertainty in the state impacts the derived stochastic approximation of the responsibility-sensitive safety constraints, which ensures desirable driving behaviors. The performance of two uncertainty-aware approaches are compared with an uncertainty-unaware baseline. The results demonstrate the effectiveness of the uncertainty-aware approach, which safely completes the driving task in 99-100% of cases while the uncertainty-unaware policy fails to do so in 42-57% cases.
Source: IFAC-PapersOnLine
Institutions: General Motors (United States), General Motors (Poland)
Language: English
Subfield: Automotive Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4406736092