Perception Without Vision for Trajectory Prediction: Ego Vehicle Dynamics as Scene Representation for Efficient Active Learning in Autonomous Driving
DOI: 10.1109/itsc60802.2025.11423425
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 high cost of data annotation in autonomous driving machine learning by proposing a method for efficient data curation using only ego vehicle dynamics. The authors argue that trajectory and dynamic state information share mutual information with visual scene observations, allowing systems to infer scene context without vision. The research aims to reduce annotation budgets while maintaining model performance by leveraging this low-cost data to guide active learning strategies, specifically focusing on trajectory prediction tasks. The methodology involves clustering trajectory-states, defined as 2D ground plane coordinates combined with vehicle state variables (velocity, acceleration, and heading change rate). The authors define a similarity metric to measure distance between trajectory-states and apply hierarchical clustering to group similar trajectories. Based on these clusters, they propose a novelty-sensitive active learning algorithm that selects data for annotation based on two parameters: $\alpha$, the proportion of novel (unrepresented) data to sample, and $\beta$, the depth of sampling within clusters. This approach allows the system to balance sampling typical data to overcome the "cold start problem" and sampling novel data to ensure diversity as the training pool grows. Experiments were conducted on the nuScenes dataset using the Prediction via Graph-based Policy (PGP) model for trajectory prediction. The authors evaluated the active learning strategy against random sampling across training pool sizes ranging from 10% to 50% of the dataset. Results demonstrate consistent performance gains over random sampling in terms of minimum average displacement error (minADE5 and minADE10). Notably, the active learning approach achieved performance comparable to or better than models trained on the full dataset using only 50% of the data. The study also identified a "phase transition" in learning strategy: sampling typical data was most beneficial at low data volumes (e.g., 10–20%), while sampling novel data became more advantageous as the data budget increased (30–50%). Qualitative analysis showed that models trained with active learning exhibited better lane conformity and multimodal prediction accuracy compared to those trained on randomly selected data. The significance of this work lies in demonstrating that efficient, robust autonomous driving systems can be developed using low-cost data curation strategies. By utilizing trajectory-state-informed active learning, the approach reduces reliance on expensive visual annotation while improving model robustness. The findings suggest that integrating novelty-sensitive sampling schedules can significantly enhance learning efficiency, offering a practical solution for managing large-scale data systems in safety-critical applications. The authors conclude that this method can be extended to other autonomous driving tasks, such as object detection and path planning, further leveraging the mutual information between vehicle dynamics and scene perception.
Key finding
A trajectory-state-informed active learning strategy consistently outperforms random sampling and achieves full-dataset performance levels using only 50% of the training data by adapting sampling strategies based on dataset size.
Methodology
dataset
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.
| Stage | Outcome | Tool | Model | Prompt | Attempts | Completed |
|---|---|---|---|---|---|---|
| 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 | 2026-06-07 |
| chunk | success | chunk | — | — | 1 | 2026-06-07 |
| embed | success | embed | Qwen/Qwen3-Embedding-8B | — | 1 | 2026-06-07 |
| enrich | success | semantic_scholar | — | — | 4 | 2026-06-15 |
| 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).
- Theoretical Contribution: computational model