EagleVision Benchmark Finally Tests Self-Driving AI at 260 km/h Racing Speeds
A unified LiDAR benchmark stitches together the Indy Autonomous Challenge, A2RL, and simulator data to test how urban-trained models handle 170 mph racing.
ProTakeaways
- EagleVision unifies Indy Autonomous Challenge, A2RL, and simulator LiDAR data into one racing perception benchmark.
- Releases 14,893 Indy, 1,163 A2RL, and 12,000 simulator frames with new 3D bounding box annotations.
- Urban pretraining lifts detection NDS from 0.69 (scratch) to 0.72 on racing data.
- Intermediate racing-to-racing pretraining hits NDS 0.726 on A2RL, beating simulator-only transfer.
- Indy-trained forecaster beats in-domain A2RL training on A2RL (FDE 0.947 vs 1.250).
- Dataset, protocol, and baselines are public at avlab.io/EagleVision.
Self-driving research has been trained almost entirely on city streets, where cars move at 30 mph and pedestrians are the main worry. Autonomous racing breaks every assumption in that setup. Vehicles close on each other at combined speeds north of 300 km/h, LiDAR returns get sparse fast, and there are no lane lines or crosswalks to anchor a model. A new benchmark called EagleVision tries to make that gap measurable rather than anecdotal.
The paper introduces a unified LiDAR benchmark for 3D detection and trajectory prediction across three domains: real Indy Autonomous Challenge runs, real Abu Dhabi Autonomous Racing League (A2RL) footage, and a high-fidelity simulator. The authors release freshly annotated 3D bounding boxes for 14,893 frames from the Indy Autonomous Challenge dataset, 1,163 frames from the A2RL Real competition, and 12,000 simulator-generated frames, all standardized under a common evaluation protocol.
Why racing breaks urban perception
The case for a benchmark like this comes down to physics. In the Indy Autonomous Challenge, modified Dallara AV-21 racecars run at the handling limits with a redundant sensor stack of LiDARs, cameras, and radars. The