Quantum Multimodal Driving Forecasts, Using 9 Qubits, Enable Accurate Trajectory Prediction for Autonomous Vehicles

Accurate prediction of future vehicle movements represents a critical challenge for autonomous driving systems, demanding both precision and computational efficiency. Navneet Singh and Shiva Raj Pokhrel, from Deakin University, address this need with a novel forecasting approach that leverages the inherent structure of road scenes. Their team develops a compact model which predicts subtle deviations from expected movements, rather than absolute positions, and operates within a lane-aligned framework. This innovative system, utilising a unique combination of quantum-inspired techniques, achieves state-of-the-art performance on a standard benchmark, demonstrating significantly improved accuracy and reliability in forecasting vehicle trajectories over a two-second horizon, and represents a substantial step towards safer and more effective autonomous navigation.

Predicting Vehicle Trajectories for Autonomous Driving

This research addresses the crucial challenge of predicting the future movements of vehicles and other agents in complex environments, a fundamental requirement for safe and reliable autonomous driving. Accurate trajectory prediction is essential for enabling autonomous vehicles to plan their movements, avoid collisions, and navigate challenging scenarios. Current methods face limitations in modeling complex interactions, predicting over long time horizons, handling uncertainty, and achieving real-time performance. Researchers are now exploring the potential of Quantum Machine Learning (QML) to overcome these limitations, offering the possibility of representing complex data more efficiently, accelerating computation, and improving model capacity. This work suggests that QML holds promise for improving the accuracy, robustness, and generalizability of trajectory prediction systems, paving the way for safer and more efficient autonomous vehicles. Beyond autonomous driving, this research has implications for robotics, logistics, human motion forecasting, and even genomic data analysis, demonstrating the broad applicability of QML techniques for sequential prediction problems.

Quantum Prediction of Vehicle Trajectories Demonstrated

Researchers have pioneered a new approach to trajectory forecasting for autonomous driving, employing a compact hybrid quantum-classical architecture designed for accurate predictions under strict computational constraints. The team formulated lane-aligned residual prediction as a quantum sequence modeling problem, constructing a low-depth hybrid quantum neural network (QNN) operating on nine-dimensional state histories of surrounding vehicles. Rigorous comparisons against classical baselines demonstrate improved accuracy and calibration. The methodology begins with extensive data preprocessing of the Waymo Open Motion Dataset, utilizing approximately 100,000 driving scenarios.

The team extracted trajectories and high-definition map data to predict the future trajectory of a self-driving vehicle based on its past motion and map context. A crucial step involved transforming each scenario into an ego-centric, lane-aligned coordinate frame, positioning the vehicle at the origin and aligning the x-axis with the local road direction. The core of the approach lies in a quantum encoder, which compresses the vehicle’s motion history into a compact representation. This encoder implements an attention-like mechanism using nine qubits, injecting classical data with single-qubit rotations. The resulting system delivers calibrated multi-hypothesis outputs under tight compute budgets, addressing a critical need in autonomous driving applications.

Quantum Trajectory Forecasting for Autonomous Driving

This work presents a novel trajectory forecasting system for autonomous driving, achieving high accuracy and calibrated multi-modal predictions within strict computational limits. The system is based on a compact hybrid architecture leveraging a 9-qubit quantum circuit, designed to align with the inherent structure of road scenes. Experiments on the Waymo Open Motion Dataset demonstrate an average displacement error of 1. 94 meters and a final displacement error of 3. 56 meters when predicting trajectories over a 2.

0-second horizon, consistently outperforming a kinematic baseline. The system predicts trajectory residuals within an ego-centric, lane-aligned frame, effectively compressing the output range and removing rigid-body motion. A truncated Fourier decoder generates 16 trajectory hypotheses in a single pass, with mode confidences derived from the latent spectrum. Training utilizes Simultaneous Perturbation Stochastic Approximation (SPSA), avoiding the need for backpropagation through non-analytic components. Detailed analysis reveals that the validation loss closely tracks the training loss, indicating generalization without overfitting, and that the shallow circuits admit stable optimization and predictable convergence. Fine-grained learning signals demonstrate steady improvement in both raw and smoothed loss over 100 epochs, coinciding with SPSA schedule resets. This performance is attributed to the system’s bias toward smooth, lane-conforming adjustments, facilitated by the truncated Fourier basis, and demonstrates a substantial advancement in accurate and reliable trajectory forecasting for autonomous vehicles.

Compact Quantum Architecture for Trajectory Forecasting

This research presents a compact hybrid quantum architecture designed for short-horizon trajectory forecasting, a critical component of autonomous driving systems. By focusing on predicting residual corrections within an ego-centric, lane-aligned framework, the model efficiently concentrates on meter-scale adjustments rather than complete trajectory predictions. The team successfully integrated a transformer-inspired quantum attention encoder, a parameter-lean feedforward network, and a Fourier-based decoder into a 9-qubit pipeline capable of generating sixteen trajectory hypotheses in a single pass. Evaluations on a representative subset of the Waymo Open Motion Dataset demonstrate that the model achieves meter-scale average and final displacement errors, consistently surpassing the performance of a strong kinematic baseline.

Furthermore, the system exhibits stable miss rates and strong precision-recall behaviour, indicating reliable multi-modal forecasting capabilities. The researchers highlight that the spectrum-based ranking method effectively captures representational headroom, allowing the model to identify plausible future trajectories. Training with the Simultaneous Perturbation Stochastic Approximation method proves stable despite the presence of non-differentiable components, demonstrating the feasibility of optimizing shallow, few-qubit circuits for this application. While the study acknowledges limitations including reliance on classical simulation, a single-vehicle formulation, and a short prediction horizon, future work aims to deploy the architecture on near-term quantum hardware, incorporate richer contextual information, and explore more expressive quantum decoders. Beyond autonomous driving, the researchers suggest that the residual, physics-shaped formulation and phase-based multi-modality techniques developed here are applicable to other resource-constrained sequential prediction problems in robotics, logistics, and human motion forecasting.