Edge for Autonomous Vehicles
Overview
Edge is an interactive tool for managing and exploring all data related to the training and testing of autonomous vehicles. Using geospatial, taxonomy, and custom layouts of data, users can intuitively understand their system’s failure modes on the micro and macro scale, leading to a faster and better prioritisation of development tasks.
Data Model
A handful of the most important types of nodes you will encounter while working with autonomous vehicle datasets in Edge are outlined below.
Scenarios
Whether ingested from real-life video and text data, or generated via synthetic methods,
Edge enables users to manage and intuitively interact with scenario data no matter the source.
From each scenario node, users can navigate to entity- and trajectory-level information for detailed
performance analysis.
Annotations
To make sense of the vast scenario landscape of real-world risk, scenarios are characterised
in Edge by the features they contain. These features, whether categorical (e.g. junction and hazard type) or numerical
(e.g. speed and relative angle of the hazard) are encoded in Edge using annotation nodes which in bulk form a Taxonomy
that can be used to navigate through scenario data in Edge. dRISK automatically annotates all scenarios with a taxonomy
of 300+ environment, trajectory, and entity level annotations.
Custom performance metrics can easily be added the Taxonomy.
Road Networks
All road networks relevant to your test suite come prepopulated in Edge. Interact with your
scenarios in geospatial views on the micro and macro scale: project individual scenarios into their spatial layouts
for analysis of entity trajectories, and project sets of scenarios onto their local ODD for analysis of
geospatial coverage and risk along a route.
LHS: The trajectories of a single scenario overlaid on the road network,
RHS: A set of scenario nodes which take place on the same intersection
Example Workflows
The capabilities of Edge for Autonomous Vehicles fall into the following three categories: Dataset Exploration, Scenario Management, and Performance Testing. Jump in using the following three workflows.
Exploring Heterogeneous Datasets from Across The Globe
Using Edge, you can import a broad variety of data sources (e.g. video, image, text, table) and fuse them under a single taxonomy. In this workflow, we explore and then query the real-life traffic incidents in the AV sandbox dataset.
empty line
Locating and Exploring Pedestrian Scenarios
In this quick workflow, we locate and browse all the scenarios which contain a pedestrian hazard and explore their taxonomic coverage. Another great place to get started!
empty line
What has gone wrong with AVs? What's going wrong next?
Here we will see examples of the high-risk scenarios autonomous vehicles have faced so far, and a prediction of what they are going to have to deal with next as deployments roll out in new cities, and hence a wider ODD.
empty line
Where to take my driver out?
Figuring out when taking the safety driver out can be a complex task with high-stakes. In the following demo, we use Edge to determine failure modes in specific areas, see where those failure modes don’t occur, and use that understanding as a place to consider a deployment without a safety driver.
empty line
The dRISK AV Sandbox
The workflows above utilise dRISK's Autonomous Vehicle sandbox dataset. This graph contains a fusion of structured and unstructured collision datasets, static video data from accident hotspots, dash-cam videos, media articles, and simulated recreations of risky events to map out the tricky events and phenomena autonomous vehicles face on the road. The dataset is accessible from your homepage and contains 10 pre-populated views for you to explore.
You can also access the AV Sandbox here: link to data Open Graph in Edge →
What does each view contain?
| View Title | Content | Description |
|---|---|---|
Example Data Sources - Spatial View |
 |
The nodes in this view represent real-world high-risk events that have taken place across the US and UK, sourced from accident databases, CCTV and dashcam videos, accident report forms and media articles. |
Example Data Sources - Taxonomy View |
 |
This view contains annotations which connect to the scenario nodes in Example Data Sources - Spatial View. |
AV failures mentioned in the media |
 |
This view contains examples of high-risk scenarios which AVs have struggled with in the past, alongside a handful of annotations that describe the events. |
Scenario Results - Taxonomy View |
 |
In this view we have provided an example set of ~600 dRISK scenarios, sourced from real-life datasets, alongside a subset of the "Full Scenario Taxonomy" for navigation. |
Scenario Spatial View |
 |
The same scenario nodes contained in "Scenario Results - Taxonomy View" laid out spatially across the map on which they take place. |
Histogram - Maximum Speed of Hazard |
 |
The same set of ~600 example scenarios are distributed in a histogram, according to the speed of the hazard vehicle involved in the scenario |
Bar Plot - Junction Types |
 |
An example bar plot which plots the example scenario set according to the junction-type around which the scenario took place. |
Feature Map - example scenarios |
 |
In this view, the scenario nodes are clusters according to the type of junction, collision, and hazard involved in the scenario. |
Scenario Embedding |
 |
This view contains a mixture of real-world video data and dRISK scenarios, laid out in an embedding such that scenarios with similar stories are clustered together. |
Ful Scenario Taxonomy |
 |
This view contains over 500 descriptive scenario annotations, such as vehicle trajectory metrics and details of the local environment in which the event took place. |
Other Autonomous Vehicle Use Cases
Scenario Selection
With Edge you can quickly find any scenarios you need from within the knowledge graph, and store them ready for upcoming simulation tests. In this workflow, we locate and organise a few example scenario types.
Scenario Generation
Sensor failure modes can be found by testing sensor software in simulation on scenarios with a variety of sensor configurations and hazardous obstacle types. Such scenarios can be easily generated using the dRISK Edge tool. In this workflow, we walk through how requirements-based scenarios can be generated in dRISK Edge.
Test Session Management
An introduction to how Edge can be used for organising scenario simulation experiments in which a certain stack, or multiple stacks, are tested against specific scenario datasets.
Use case come to mind, but not listed on this page? Reach out to shona@drisk.ai - whatever it is, Edge will speed it up.