Towards Scalable & Efficient Interaction-Aware Planning in Autonomous Vehicles using Knowledge Distillation

TL;DR

This work tackles the computational bottleneck of integrating neural network based interaction predictions with MPC for autonomous vehicle planning. It introduces a teacher–student knowledge distillation framework to train a compact student network that performs one shot interactive trajectory predictions over a horizon $t_{plan}$, replacing iterative teacher in the planning loop. Empirical results show the student achieves about a sixfold reduction in state update time and enables planning at 2–4 Hz with similar optimization costs to the teacher, illustrating real time viability without notable loss in accuracy. The approach broadens the applicability of neural network–driven constrained optimization in robotics and autonomous driving by enhancing scalability and efficiency while maintaining safety and performance guarantees.

Abstract

Real-world driving involves intricate interactions among vehicles navigating through dense traffic scenarios. Recent research focuses on enhancing the interaction awareness of autonomous vehicles to leverage these interactions in decision-making. These interaction-aware planners rely on neural-network-based prediction models to capture inter-vehicle interactions, aiming to integrate these predictions with traditional control techniques such as Model Predictive Control. However, this integration of deep learning-based models with traditional control paradigms often results in computationally demanding optimization problems, relying on heuristic methods. This study introduces a principled and efficient method for combining deep learning with constrained optimization, employing knowledge distillation to train smaller and more efficient networks, thereby mitigating complexity. We demonstrate that these refined networks maintain the problem-solving efficacy of larger models while significantly accelerating optimization. Specifically, in the domain of interaction-aware trajectory planning for autonomous vehicles, we illustrate that training a smaller prediction network using knowledge distillation speeds up optimization without sacrificing accuracy.

Figures (10)

• Figure 1: An illustrative scenario demonstrating the impact of interaction awareness. In (a), the interaction-unaware AV (orange) waits for an extended period, while in (b), the interaction-aware AV (orange) predicts that signaling or nudging left will prompt the blue vehicle behind the AV to decelerate, creating an opportune gap for a successful left merge.
• Figure 2: Neural-network-based optimization integrates a neural network with closed-loop control, utilizing the neural network's output for optimization and feeding the optimization output back into the neural network.
• Figure 3: Overview of the Social-GAN's RNN based Encoder-Decoder generator for a one time-step prediction ($t_{pred} = 1$).
• Figure 4: Overview of the interactive trajectory generator for the (a) teacher network and the (b) student network. (a) The teacher network utilizes Social-GAN's generator multiple times to generate interactive trajectories for $t_{plan}$ time steps. (b) The student network produces the complete interactive trajectory by taking the ego candidate trajectory as input and running the decoder for $t_{plan}$ time steps.
• Figure 5: Student network outputs the interactive trajectory for $t_{plan}$ time-steps by using the ego candidate trajectory as the input.

Theorems & Definitions (3)

• Remark 1
• Remark 2
• Remark 3