Computational Teaching for Driving via Multi-Task Imitation Learning

Gopinath, Deepak; Cui, Xiongyi; DeCastro, Jonathan; Sumner, Emily; Costa, Jean; Yasuda, Hiroshi; Morgan, Allison; Dees, Laporsha; Chau, Sheryl; Leonard, John J.; Chen, Tiffany; Rosman, Guy; Balachandran, Avinash · 2025 · Unknown

DOI: 10.1109/icra55743.2025.11127621

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 creating automated teaching systems for complex motor skills, specifically high-performance track driving. While human instructors provide effective, context-aware verbal guidance, their availability is limited and costly. Training automated systems via imitation learning is hindered by the scarcity of high-quality, annotated datasets capturing expert-student interactions. To overcome this data bottleneck, the authors propose a Multi-Task Imitation Learning (MTIL) framework that leverages abundant, non-interactive driving data to learn robust representations, thereby improving the prediction of teaching instructions using only small amounts of labeled interactive data. The proposed model employs an encoder-decoder architecture that processes a student’s past driving trajectory and local map information. The encoder fuses these inputs using cross-attention mechanisms, while the decoder predicts three outputs: teacher actions (primary task), future vehicle trajectories, and driver skill estimates (auxiliary tasks). The system is trained using a combined loss function that includes weighted binary cross-entropy for action classification, minimum-over-N average displacement error for trajectory prediction, and mean squared error for skill estimation. This design allows the model to utilize self-supervised signals from unlabeled driving data to enhance the primary task of instruction generation. The approach was validated through four distinct experiments. First, on a semi-synthetic dataset derived from the Waymo Open Motion Dataset, results demonstrated that adding auxiliary tasks significantly improved teacher action prediction accuracy, particularly when the teaching logic relied on inferred driver skill levels. Second, on a novel dataset of real-world professional track driving instructions, the MTIL model achieved a weighted F1-score of 77.8%, outperforming single-task baselines and maintaining performance even when trained on only 20% of the labeled data. Third, a human-in-the-loop study using a high-fidelity driving simulator showed that students receiving automated instructions improved their lap times more rapidly in early learning phases and spent less time outside track bounds compared to a control group. Finally, the system was successfully deployed on an instrumented Lexus LC500, demonstrating real-time capability on physical hardware. The study concludes that multi-task learning with self-supervised auxiliary tasks is an effective strategy for computational teaching in domains where interactive training data is scarce. By jointly learning trajectory prediction and skill estimation, the model acquires latent features that enhance the quality of generated instructions. The findings suggest that automated coaching systems can effectively support human learning in complex sensorimotor tasks, offering a scalable alternative to one-on-one human instruction. This work opens new avenues for intelligent tutoring systems in robotics and human-robot interaction, particularly where expert supervision is resource-intensive.

Key finding

Integrating self-supervised auxiliary tasks like trajectory prediction and skill estimation into an imitation learning framework significantly enhances the accuracy of automated driving instruction predictions and improves student driving performance in simulator environments.

Methodology

mixed_methods

Sample size: 15

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. Discovered via author_sweep_intake on 2026-05-28.

StageOutcomeToolModelPromptAttemptsCompleted
discover success author_sweep 2 2026-05-28
archive success canonical_url 1 2026-06-06
extract success cached 3 2026-06-10
clean success clean 1 2026-06-04
chunk success chunk 1 2026-06-04
embed success embed Qwen/Qwen3-Embedding-8B 1 2026-06-04
enrich success 1 2026-05-28
promote success 1 2026-06-04
summarize success llm qwen3.6-27b-prismaquant summ-v5 2 2026-06-10
tag success vector_similarity 15 2026-06-11
verify success 2 2026-06-10

Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.

Topics

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

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).