Inverse Learning for Data-Driven Calibration of Model-Based Statistical Path Planning

Abstract

This paper presents a method for inverse learning of a control objective defined in terms of requirements and their joint probability distribution from data. The probability distribution characterizes tolerated deviations from the deterministic requirements and is learned using maximum likelihood estimation from data. Further, this paper introduces both parametrized requirements for motion planning in autonomous driving applications and methods for the estimation of their parameters from driving data. Both the parametrized requirements and their joint probability distributions are estimated using a posterior distribution such that the control objective is personalized from a prior as driver data are accumulated. Finally, three variants of the learning method are presented that vary in computational complexity and data storage requirements. Key advantages of the proposed inverse learning method are a relatively low computational complexity, a need for a limited amount of data, and that the data do not have to be segmented into specific maneuvers, which makes the method easily implementable. Learning results using data of five human drivers in a simulation environment suggest that the proposed model for human-conscious driving along with the proposed learning method enable a more natural and personalized driving style of autonomous vehicles for their human passengers.

Interaction-Aware Motion Prediction for Autonomous Driving: A Multiple Model Kalman Filtering Scheme

“…where w LQR for each driver. The individuality of different drivers is highlighted, e.g., in [8], [22].…”

“…Proof: Without loss of generality, let the priority list l in (21) be sorted in ascending order according to the vehicles' indices. Then, the index sets in (22) take the form…”

Interaction-Aware Trajectory Prediction and Planning for Autonomous Vehicles in Forced Merge Scenarios

“…Then we can utilize (19) to update the belief of each interacting vehicle's leader or follower role, P(σ k = l|ξ k t ), l ∈ L = {leader, follower}. The MPC-based control strategy presented in ( 6) can be reformulated as…”

“…The decision making algorithm proceeds as follows: At the sampling time t, the ego vehicle measures the current states of each pairwise interaction and adds them together with the previous control input to the observation vectors ξ k t . The belief about each vehicle's leader or follower role is updated according to (19) based on ξ k t . Then, the MPC-based control strategy ( 22) is utilized to obtain the optimal trajectory (γ 0 ) * by searching through all trajectories introduced in Section II-D, and the ego vehicle applies the first control input (u 0 t ) * over one sampling period to update its states.…”

“…An RL-based policy can account for the vehicle interactions in certain scenarios through training in an environment capable of representing such interactions [14], [15]. In order to obtain through RL driving policies that behave like human drivers, several researchers chose to use inverse RL to estimate human's reward function for driving [16], [17], [18], [19]. To be able to model different human driver styles and/or interaction intentions, [20] incorporates cooperativeness into the intelligent driver model and [21] formulates different reward functions for different drivers and performs RL based on the models.…”