Research map / Label-efficient and robust learning / Knowledge distillation

On the Road to Portability: Compressing End-to-End Motion Planner for Autonomous Driving

CVPR 2024 · Label-efficient and robust learning › Knowledge distillation · approach: Patient consistent distillation · 16 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes PlanKD, a knowledge distillation framework that compresses end-to-end CARLA motion planners (InterFuser, TCP) using information-bottleneck planning-relevant feature distillation and safety-aware waypoint-attentive distillation.

Problem

End-to-end motion planners for autonomous driving are too large for resource-constrained deployment, and naive smaller backbones degrade driving performance sharply; standard knowledge distillation does not account for planning-relevant information or waypoint safety importance.

Approach

Distill only planning-relevant intermediate features via a variational information bottleneck tied to eight planning states, and weight waypoint imitation by a safety-aware attention trained with a ranking loss that favors waypoints close to moving obstacles.

Key results

• On Town05 Long, a 26.3M InterFuser student with PlanKD reaches a driving score of 55.90 versus 36.55 without KD (teacher 53.44), at about 39.7 ms inference time.
• A 13.9M TCP student with PlanKD reaches 53.19 driving score on Town05 Long versus 39.96 without PlanKD (teacher 53.41).
• On Town05 Short, PlanKD outperforms AT, ReviewKD and DPK; the 26.3M InterFuser student reaches 93.69 driving score vs 84.59 for the best baseline DPK.

Measured effects reported in the paper

Surprising findings

• Route Completion of PlanKD students is sometimes lower than other KD methods, which the authors attribute to safety-prioritizing stop decisions.
• Some PlanKD students (e.g., 26.3M InterFuser on Town05 Long) outperform their teacher's driving score despite being half the size.

Limitations

• Evaluated only in the CARLA simulator on Town05 benchmarks
• Route Completion can be lower than some KD baselines because the planner prioritizes stopping for safety
• Only two teacher architectures (InterFuser and TCP) are compressed

Experimental setup

Datasets: CARLA simulator data (800K frames, 7 towns for training)

Benchmarks: Town05 Long Benchmark Town05 Short Benchmark

Metrics: Driving Score Route Completion Infraction Score Collision Rate Infraction Rate inference time

Baselines: InterFuser (reduced size, no KD) TCP (reduced size, no KD) AT ReviewKD DPK

Builds on: InterFuser TCP information bottleneck attention transfer (AT) ReviewKD DPK

Closest papers by meaning

• LaKD: Length-agnostic Knowledge Distillation for Trajectory Prediction with Any Length Observations (NeurIPS 2024 · similarity 0.87)
• Drive-R1: Bridging Reasoning and Planning in VLMs for Autonomous Driving with Reinforcement Learning (AAAI 2026 · similarity 0.86)
• Don't Shake the Wheel: Momentum-Aware Planning in End-to-End Autonomous Driving (CVPR 2025 · similarity 0.86)
• Enhancing End-to-End Autonomous Driving with Latent World Model (ICLR 2025 · similarity 0.85)
• UniMotion: A Unified Motion Framework for Simulation, Prediction and Planning (NeurIPS 2025 · similarity 0.85)
• ChainFlow-VLA: Causal Flow Planning with Vision-Language Models (NeurIPS 2026 · similarity 0.85)