Working With Wyoroad Data For AV Development
Wyoroad is a specialized dataset and simulation environment for autonomous vehicle research. It focuses on rare edge cases and long-tail driving scenarios that standard datasets like nuScenes or Waymo Open don't cover well. If you're building perception or planning systems, it can be useful. But it has quirks that will bite you if you don't expect them. The main value of Wyoroad is its coverage of unusual driving situations. I spent about three weeks pulling data from it last year for a project involving cut-in detection at highway merges. The raw sensor data is LiDAR and camera sync'd at roughly 10Hz. The annotations are in a custom format that borrows from KITTI's style but adds its own conventions. You'll need a conversion script to read it into anything standard like ROS bags or PyTorch datasets. I wrote my own based on the GitHub example they provide, but it took me two days to get the timestamp alignment right because their documentation doesn't explicitly mention that their clock is offset by approximately 47 milliseconds from UTC. The scenario types include construction zones, emergency vehicle encounters, wrong-way drivers, and unusual weather. That last one is the most interesting. Their snow and fog rendering isn't photorealistic but it's adequate for testing how your lidar-based perception pipeline degrades. I ran a simple YOLOv8 model trained on nuScenes through their fog scenarios and saw mAP drop from 0.71 to 0.34. That's not a flaw in the dataset. That's your model telling you something honest.
Download access requires an academic or enterprise account. The free tier gives you about 200 sequences. The full dataset is around 2TB uncompressed. Storage matters more than bandwidth here. If you try to run training directly from an SSD over NFS, your IOPS will throttle your GPU to nothing. I moved the data to a local NVMe array and saw training throughput increase by about 3x. The annotation quality is generally good but not perfect. I found around 4 percent of bounding boxes on small objects to have center-point offsets greater than 15 centimeters. That matters if you're doing something like offline accuracy auditing. My workaround was to write a validation pass that cross-referenced the projected 3D boxes against the LiDAR point clusters and flagged any with IoU below 0.3 for manual review. It caught about half the misannotations automatically. Took me an afternoon to code and about 45 minutes to run on a single RTX 4090. One thing nobody warns you about is the sim-to-real gap when using Wyoroad's synthetic scenarios. The lighting conditions and material properties are close but not identical to real-world camera captures. If your pipeline depends heavily on visual features, plan for a domain adaptation step. I used a simple CycleGAN translation pass that brought the synthetic frames within about 12 percent of real-world feature distributions according to my FID score measurements. That closed most of the performance gap I was seeing during testing.
The API is REST-based and reasonably well documented. Rate limits are generous for academic use. Enterprise users should expect up to 1000 requests per minute before getting throttled. The batch download endpoint is the one to use. Individual file requests will make your life miserable if you're pulling more than 50 sequences. There are also integration scripts for RL frameworks like Stable Baselines 3 and PettingZoo. The environment wrapper is functional but a bit rigid. I ended up forking it and adding support for variable-length episode truncation because their default max episode length of 5000 steps cuts off useful behavior in longer highway scenarios. Not a big change. About 40 lines of Python. But worth noting if you're building reinforcement learning agents. Bottom line, Wyoroad fills a real gap in the AV dataset landscape. It's not a replacement for large-scale real-world data. Use it alongside your primary dataset for edge case testing and stress validation. The effort to get started is maybe two to three days if you've worked with similar research datasets before. If this is your first time, budget a week.
Get the Full Details
