Decision-Making for Autonomous Vehicles with Interaction-Aware Behavioral Prediction and Social-Attention Neural Network
ArXived:
Xiao Li, Kaiwen Liu, H Eric Tseng, Anouck Girard, Ilya Kolmanovsky
Autonomous vehicles, as players in real-world traffic, need to accomplish their tasks while interacting with human drivers. It’s crucial to equip the autonomous vehicle with artificial reasoning to better comprehend the intentions of the surrounding traffic, thereby facilitating their accomplishments of the tasks. In this work, we proposed a behavioral model that encodes drivers’ interacting intentions into latent social-psychological parameters. Leveraging a Bayesian filter, we develop a receding-horizon optimization-based controller for the autonomous vehicle’s decision-making, while counting for the uncertainties in interacting drivers’ intentions. For online deployment, we design a neural network architecture based on the Attention mechanism that imitates the behavioral model with online estimated parameter priors. We propose a decision tree search algorithm to online solve the decision-making problem. The behavioral model is evaluated in the task of real-world trajectory prediction. We further conduct extensive evaluations of the proposed decision-making module, in forced highway merging scenarios, using both simulated environments and real-world traffic datasets. The results demonstrate our algorithms can complete the forced merging tasks in various traffic conditions while ensuring driving safety. [ArXiv]
Proactive Forced Merging in Simulation: Our ego vehicle proactively interacts (i.e., tentative merging action) with the “egoistic” driver 1 to test the driver’s intention. Our algorithm can effectively interpret the “egoistic” identity of driver 1. Moreover, our decision-making module can leverage this information to facilitate the forced merging task while ensuring the safety of the ego vehicle.
Forced Merging in the Real-World High-D Dataset: We evaluate the performance of our method in the High-D real-world traffic dataset. There are 60 recordings in the High-D dataset where the recording 58-60 corresponds to highways with ramps. In recording 58-60, we identify in total 75 on-ramp vehicles (High-D target vehicles) that merge into highways. The following animations show some of the test cases.
Forced Merging in the Carla Simulator: We evaluate the performance of our method in the Carla simulator with various of traffic densities and drivers’ probability of yielding. The following animations show some of the test cases.
