Autonomous vehicle technologies have seen a surge in popularity over the last decade, with the potential to improve flow of traffic, safety and fuel efficiency . In any traffic scenario, an autonomous agent must plan to follow its desired route while accounting for surrounding agents to maintain safety. The difficulty arises from the variability in the possible behaviors of the surrounding agents. To address this difficulty, significant research has been devoted to modelling these agent predictions as multi-modal distributions. Such models capture uncertainty in both high-level decisions (desired route) and low-level executions (agent position, heading, speed). Motion planning algorithms for autonomous vehicles using these multi-modal predictions, require finding the right balance between performance, safety and computational tractability. Furthermore, our recent research addresses the scalability of the MPC-based path planning with the number of vehicles and their multi-modal predictions leveraging duality-based interaction predictions.
On this page, we highlight our work on planning and predictions using multi-modal distributions for autonomous driving.
We propose a hierarchical architecture designed for scalable real-time Model Predictive Control (MPC) in complex, multi-modal traffic scenarios. This architecture comprises two key components: 1) RAID-Net, a novel attention-based Recurrent Neural Network that predicts relevant interactions along the MPC prediction horizon between the autonomous vehicle and the surrounding vehicles using Lagrangian duality, and 2) a reduced Stochastic MPC problem that eliminates irrelevant collision avoidance constraints, enhancing computational efficiency. Our approach is demonstrated in a simulated traffic intersection with interactive surrounding vehicles, showcasing a 12x speed-up in solving the motion planning problem. A video demonstrating the proposed architecture in multiple complex traffic scenarios can be found here
Baseline: Full MPC with collision avoidance constraints imposed for all multi-modal predictions
Proposed: Sparse MPC with fewer collision avoidance constraints based on duality-based interaction predictions
Publications:
Github repository: Code
Planning in open-loop
Planning using feedback policies
We propose Stochastic MPC formulations that optimize over feedback policies designed to exploit the structure of the multi-modal predictions, and are amenable to convex programming. The use of feedback policies for prediction is motivated by the need for reduced conservatism in handling multi-modal predictions of the surrounding vehicles. Our algorithm shows considerable benefits compared to prevailing approaches along axes of mobility, comfort, conservatism and maintains computational efficiency.
Multi-modal Prediction Architecture
Despite the benefits of context-aware, data-driven, multimodal predictions, the errors in prediction during online planning may persist. This is attributed to the prediction models learning rational, normative driving behavior. We propose an adaptive scheme to capture the consistency of the prediction model using confidence thresholds, updated in a Bayesian fashion.
Thesis:
Github repository: conf_aware_preds