BucketKD: A Safety-Aware Bucket-Based Knowledge Distillation Framework for End-to-End Motion Planning

Jul 16, 2026

Upload by Miles

End-to-end motion planning directly maps raw sensor inputs to driving control through deep neural networks, but the approach is often blocked by large model size that limits deployment on resource-constrained platforms. BucketKD introduces bucket-based knowledge distillation that compresses the planner while remaining safety-aware. It discretizes critical environmental variables into adaptive, nonuniform buckets for richer scene semantics. A safety-aware waypoint attention mechanism uses time-to-collision risk and relative motion to preserve safety-critical behaviors. Experiments on CARLA with Bench2Drive show improved planning accuracy and safety with strong compression.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages8

Reading Time20 min

Introduction

End-to-end motion planning overview

Model compression via knowledge distillation

Limitations of prior approaches

Proposed method: BucketKD

Bucket-based planning state formulation

Safety-aware waypoint attention with TTC

Experimental evaluation

BucketKD: A Safety-Aware Bucket-Based Knowledge Distillation Framework for End-to-End Motion Planning Md Nahidul Islam2 , Mohd Hasan Ali1 , Dipankar Dasgupta2 , and Myounggyu Won2 arXiv :2607 . 10565v1 [ cs .RO] 12 Jul 2026 Abstract—End-to-end motion planning has emerged as a promising paradigm in autonomous driving, directly mapping raw sensor data to control commands via deep neural networks. Despite its advantages, its large model size hinders deployment in resource-constrained platforms. In this paper, we present BucketKD, a bucket-based knowledge distillation framework that yields compact and safety-aware end-to-end planners. Compared to the state-of-the-art approach, which relies onsimpliﬁed planning state representations, BucketKD discretizes critical environmental variables into adaptive buckets that capture richer scene semantics while preserving efﬁciency. In addition, we design a safety-aware waypoint attention mechanism that evaluates each waypoint’s risk level by accounting for both obstacle proximity and relative motion through a timeto-collision (TTC) formulation widely used in transportation research. This enables the student model to better retain safetycritical behaviors during distillation. Extensive experiments in CARLA using the Bench2Drive dataset show that BucketKDsigniﬁcantly outperforms the state-of-the-art in both planning accuracy and safety while maintaining strong compression ratios. I. INTRODUCTION End-to-end motion planning has emerged as a promising paradigm in autonomous driving. In this approach, deep neural networks learn an end-to-end mapping from raw sensor data to the vehicle’s trajectory or control actions, removing the need for explicit perception, planning, and control submodules [1], [2], [3], [4]. End-to-end motion planning offers reduced latency, greater efﬁciency, and stronger generalization than traditional modular pipelines by integrating perception, prediction, and planning into a single data-driven model [5], [6] . However, a key challenge of this approach is its large model size, which typically demands highperformance onboard hardware, making deployment difﬁcult in resource-constrained environments such as autonomous delivery robots or vehicles with limited computational capacity [2] . To mitigate the model size limitation, a state-of-theart compression framework PlanKD leverages knowledge distillation to achieve compact yet high-performing endto-end motion planners [2] . Inspired by the bottleneck principle [7], PlanKD learns compact representations that 1Mohd Hasan Ali is with the Department of Electrical and Computer Engineering, University of Memphis, Memphis, TN, United States [mhali@memphis.edu](mhali@memphis.edu) 2Md Nahidul Islam, Dipankar Dasgupta, and Myounggyu Won are with the Department of Computer Science, University of Memphis, Memphis, TN, United States {mislam19, ddasgupt, [mwon](mwon}@memphis.edu)[}](mwon}@memphis.edu)[@memphis.edu](mwon}@memphis.edu) capture planning-critical features with minimal dimensionality while maintaining sufﬁcient information for robust decision-making. While PlanKD achieves substantial model compression without notable performance degradation, the oversimpliﬁcation of its planning state representation constrains its capability to handle complex and dynamic driving scenarios. Furthermore, its safety model’s dependence on distance-based metrics alone limits its ability to assess the full contextual risk of the environment, thereby increasing the likelihood of suboptimal or unsafe maneuvers. In this paper, we present BucketKD, a Bucket-based Knowledge Distillation framework for end-to-end motion planning in autonomous driving. Given the limited generalization of binary state representations in existing knowledge distillation frameworks, especially under complex and dynamically evolving trafﬁc conditions, we introduce a bucketbased method that discretizes each critical state variable into a set of nonuniform buckets

FAQs

What problem does BucketKD address in end-to-end motion planning?

BucketKD targets the large-model requirement of end-to-end planners that makes deployment difficult on limited hardware, while also improving safety during distillation.

How does BucketKD improve planning-state representation compared with prior distillation methods?

It discretizes critical environmental variables into adaptive nonuniform buckets, capturing finer-grained behavioral variations without making the state space unmanageable.

What makes BucketKD safety-aware during knowledge distillation?

It uses a safety-aware waypoint attention mechanism that assigns risk based on obstacle proximity and relative motion through a time-to-collision (TTC) formulation, helping the student retain safety-critical behaviors.

Jul 16, 2026

Upload by Miles