EagleVision: A Multi-Task Benchmark for Cross-Domain Perception in High-Speed Autonomous Racing
EagleVision: A Multi-Task Benchmark for Cross-Domain Perception in High-Speed Autonomous Racing
Authors: Zakhar Yagudin, Murad Mebrahtu, Ren Jin, Jiaqi Huang, Yujia Yue, Dzmitry Tsetserukou, Jorge Dias, Majid Khonji Date: 2026-04-13 Paper ID: openalex:2604.11400
Summary
EagleVision is a new multi-task benchmark designed to address perception challenges in high-speed autonomous racing, focusing on 3D object detection and trajectory prediction. The benchmark consolidates diverse data sources, including real-world racing and simulator environments, under a standardized evaluation protocol to assess cross-domain generalization. Empirical results reveal that racing-specific pretraining significantly outperforms urban-only models in high-dynamic scenarios. Furthermore, the findings suggest that motion-distribution coverage is a critical factor for successful trajectory prediction in cross-domain transfer settings.
Key Contributions
- Introduces EagleVision, a unified LiDAR-based multi-task benchmark for 3D detection and trajectory prediction in high-speed autonomous racing.
- Provides 14,893 frames of newly annotated Indy Autonomous Challenge data, 1,163 A2RL frames, and 12,000 simulator-generated frames.
- Demonstrates that intermediate pretraining on real racing data provides superior transfer to new racing domains compared to simulator-only adaptation.
- Shows that trajectory prediction performance benefits significantly from broad motion-distribution coverage, with Indy-trained models achieving lower FDE (0.947 vs 1.250) on A2RL test sequences.
Open Questions & Future Work
Archivist Review
I have approved the two primary racing datasets, as they represent foundational benchmarks for the specific niche of high-speed autonomous perception. I also approved the open question regarding high-speed perception generalization because it addresses the identified bottleneck of data scarcity and sim-to-real transfer in dynamic environments. EagleVision itself was rejected as a concept because it functions primarily as a benchmark suite rather than a reusable methodological mechanism.
Approved Open Questions
- Generalizing High-Speed Racing Perception: This question highlights the critical bottleneck in high-speed perception, where limited real-world data and sim-to-real gaps hinder model generalization.
Rejected Candidates
- [concept] EagleVision (
eaglevision) - not_reusable: EagleVision is a specific benchmark dataset/tooling collection rather than a distinct algorithmic concept or architectural component.
Datasets
Links
Metadata & Links
- url
- https://arxiv.org/abs/2604.11400
- paper_id
- 2604.11400
- paper_source
- openalex
- domain
- computer-vision
- tags
-
- architectures
-
- datasets
- Indy Autonomous ChallengeA2RL Real competition dataset
- concept_slugs
-
- dataset_slugs
- indy-autonomous-challengea2rl-real-competition-dataset
- skill
- TimeSeriesSkill
- created_at
- 2026-04-16T05:47:35Z