Scalable Multi-modal Model Predictive Control via Duality-based Interaction Predictions

TL;DR

This paper tackles real-time motion planning in complex, multi-modal traffic by integrating a duality-based interaction predictor with a reduced stochastic MPC. The key idea is to screen which AV–TV interactions are relevant along the horizon using RAID-Net, an attention-based recurrent network that leverages Lagrangian duality to identify active constraints, and then solve a reduced MPC that enforces only the necessary collision-avoidance constraints. Empirically, the approach yields up to a 12× speed-up over full MPC on a traffic-intersection scenario, while maintaining high safety with a low false-negative rate and a manageable number of active constraints. The work demonstrates practical impact for scalable autonomous driving in interactive environments and outlines avenues for improved guarantees, learning-from-demonstrations refinements, and generalization to other topologies.

Abstract

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: https://youtu.be/-pRiOnPb9_c. GitHub: https://github.com/MPC-Berkeley/hmpc_raidnet

Figures (4)

• Figure 1: Our hierarchical architecture for motion planning with duality-based interaction prediction. Given the environment observation, we classify which vehicles and their maneuvers can be eliminated/screened for solving a reduced, real-time MPC problem.
• Figure 2: Schematic of our Recurrent Attention for Interaction Duals Network (RAID-Net) for predicting relevant constraints for the MPC optimization problem. RAID-net is invariant to the number and order of target vehicles, and has a MPC horizon-independent memory footprint because of its recurrent architecture.
• Figure 3: An example scene in the custom unsignalized intersection environment. The autonomous vehicle (green rectangle) is approaching the intersection and interacts with the target vehicles (blue rectangles).The active and inactive constraint predictions from the $\pi^{\text{RAIDN}}$ are depicted as yellow and magenta ellipses, respectively.
• Figure 4: RAID-Net evaluation results on a test dataset: (a) histogram of the normalized loss value of RAID-Net versus MLP neural network, (b) normalized confusion matrix of RAID-Net