Interaction-Aware Trajectory Planning for Autonomous Vehicles with Analytic Integration of Neural Networks into Model Predictive Control

Gupta, Piyush; Isele, David; Lee, Donggun; Bae, Sangjae · 2023 · Crossref

DOI: 10.1109/icra48891.2023.10160890

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of conservative motion planning in autonomous vehicles (AVs), which often results in reduced traffic throughput and poor performance in dense, interactive driving scenarios. Current AV planners typically decouple prediction and planning, treating other vehicles as static obstacles or bounded disturbances. This approach fails to capture inter-vehicle interactions, forcing AVs to wait excessively for gaps rather than negotiating with surrounding agents. The authors propose an interaction-aware motion planner that integrates a neural network-based trajectory predictor directly into a Model Predictive Control (MPC) framework. This allows the AV to predict how surrounding vehicles will react to its own maneuvers, enabling complex, locally optimal actions such as nudging other cars to change lanes or speed. The method formulates the motion planning problem as a non-convex optimization task that incorporates non-linear system dynamics and neural network predictions within safety constraints. To solve this, the authors employ the Alternating Direction Method of Multipliers (ADMM). Unlike prior heuristic approaches that sample finite trajectory candidates, this method provides a formal mathematical solution with provable optimality. The authors establish sufficient conditions for the neural network—specifically regarding bounded outputs, bounded gradients, and Lipschitz differentiability—to guarantee that the ADMM algorithm converges to a local optimum. The system uses a linearized kinematic bicycle model for dynamics and a single-circle model for collision avoidance, updating predictions recursively based on the ego vehicle’s planned trajectory. Experimental results compare the proposed ADMM-NNMPC method against a baseline heuristic method (NNMPC) in two-lane and three-lane dense traffic scenarios. In the two-lane scenario, the baseline method failed to merge after 17 time steps, whereas the proposed method successfully merged in 9 steps. In the three-lane scenario, the proposed method merged in 7 steps compared to 29 for the baseline. The ADMM-NNMPC also achieved lower maximum cost values (62 vs. 80.9 in two-lane; 53.3 vs. 114.6 in three-lane) and maintained higher minimum safety distances (2.41m and 2.66m respectively) compared to the baseline. These findings demonstrate that analytically integrating neural networks into MPC allows AVs to perform safer, more efficient maneuvers by actively interacting with surrounding traffic rather than passively waiting.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• anticipation

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Theoretical Contribution: computational model