1 Department of Transportation Engineering, Tongji University, Shanghai, China
2 Key Laboratory of Road and Traffic Engineering, Ministry of Education, Shanghai, China
Existing AVs struggle to comprehend and apply common HV social norms, especially the driving skills exhibited by adept human drivers in ambiguous right-of-way scenarios. In this study, we put forth a novel framework to leverage expert priors for proactive-aware decision-making in ambiguous right-of-way, merging data-driven reinforcement learning (RL) with parameterized modeling. The proposed method is intensively validated in different driving tasks with unprotected turning scenarios, the results demonstrate that the AV can accelerate the convergence of the interaction by consistent probing and decision updates. We present demos in the following sections.
Interactive Performance in Unprotected-turning tasks
We performed a qualitative and quantitative examination of the agent’s probing process in the context of a typical unprotected left-turn interaction. The initial state of the two vehicles was based on the ambiguous right-of-way scenario obtained in Section II.B.
Interactive Performance of The Proposed StrategyInteractive Performance of The Baseline Strategy (PPO)
Interaction with HVs of Different Driving Styles
It’s worth noting that when faced with aggressive HVs, AV demonstrates a lower likelihood of choosing to probe and gaining right-of-way advantages. When facing relatively conservative interaction objects, AV has a higher probability of completing the turn ahead of time.
Interaction with Aggressive HVsInteraction with Conservative HVs
Interaction in Facing Different Ambiguities of Right-of-way
In facing different ambiguities of right-of-way, AV with the proposed method can take distinct actions, highlighting the importance of autonomous generation of this active driving behavior instead of reliance on manually coded policies.
Interaction in Ambiguous Right-of-way.Interaction in Clear Right-of-way.
Interaction Performance in Generalization Scenarios (inD dataset)
Our method exhibits robust stability across diverse scenarios, showcasing the application of our method in varied and challenging environments.
Interaction Performance in Generalization Scenarios: Scene #2 in inD dataset.Interaction Performance in Generalization Scenarios: Scene #4 in inD dataset.
Human-in-loop experiments
With our strategy, AV can interact reasonably with human-driven through vehicles, demonstrating proper safety and anthropomorphism.
The AV controlled by the proposed strategy from the view of human-driver (obtains the right of way). The AV controlled by the proposed strategy from the view of human-driver (lost the right of way).
