Yorkshire Validation

Our Story

It's long been known that many of the verification techniques required to develop driver assistance, and full self-driving systems, can be borrowed from existing safety-critical industries, such as locomotive and aerospace. Formal standards concerning these issues, such as ISO26262, have already been published for ensuring no systematic programming errors will occur in the automotive software.

Despite the existence of these verification methods, algorithmic validation is a different story.

Depicted the 'Safety Of The Intended Functionality (SOTIF)', it regards asserting that the models will behave safely, correctly, and predictably in all driving scenarios and conditions.

Humans make good drivers due to a natural ability to succeed in this area — their instinct can adapt to unexpected and unfamiliar circumstances on the road.

While self-driving 'agents' (the part responsible for the decision making), and neural-net modules, such as object detection, can be validated through extensive (and expensive) testing on real roads, these components can also partially rely on assurances drawn from simulated environments — drastically reducing the amount of journey time or test mileage required to guarantee the safety of the system.

Some simulators have attempted to fill this void. However, testing a finite number of manually created scenarios is but a small subset of the chaos which may ensue in the real world.

For this reason we envisaged a new simulator, which naturally generates the interactions between independent vehicles and other entities on real geographical roads.

PlaceSim Objective

At Yorkshire Validation we believe that prior to acceptance for use by the public, a Level 3-5 autonomous system should satisfy the following requirement:

‘There shall be no collisions for (at least) X years in location Y due to the fault of the autonomous agent.’

This would also show that the self-driving system is at least as competent as the best human drivers, since the number of collisions caused by humans is more than 0 for any sufficiently large X years, in location Y.

Features

Target Friendly

Based in C/C++, self-driving agents validated in PlaceSim can be deployed directly onto the target autonomous systems, without having to convert and re-verify the algorithmic models.

Playback Independent

Simulation results are render independent, meaning troublesome scenarios & incidents can be inspected in any graphics engine such as Blender, Unreal Engine, or Unity; thereby allowing agent development in more user-friendly environments, such as webGL, mobile or Virtual Reality.

Debug & Profile Ready

Providing support for multiple cluster implementations carries various challenges regarding the pinpointing of potential errors in a multi-core simulation, considering terminal access, process faults and the vast amount of information involved in long-running simulations (which may be of use when debugging).

PlaceSim overcomes these issues with a robust debugging system which supports multiple output formats, along with filters for isolating logging information from specific cores, vehicles (and internal classes such as the SatNav), message types and frames at or beyond a certain point in the simulation, to avoid costly logging for the entire duration.

Being designed from the ground up with efficiency in mind, PlaceSim also provides timing mechanisms for determining the performance of self-driving agents (which currently sits around 35μs — orders of magnitude faster than a typical 16.67ms or 8.33ms frame, if computing 60 or 120 times per second).

Pre-Stocked & Expandable

Included with the simulator are various tools for getting started with the platform.

A SatNav module allows for route finding within the simulated region, and provides street information to the self-driving agent.
An initial agent, agent 'alpha', naively drives to its given destination at the speed limit of the road, with no concern for other vehicles. This results in hundreds of collisions being detected, which stands as a good baseline test to develop safe driving behaviour upon this.
PlaceSim supports any vehicle, assuming a 3D model and data profile are provided; and various sensor models, imitating those employed on an autonomous vehicle.
A grid splitting tool is included for producing large simulation regions, such as cities, to reduce the system load when reconstructing the output.
Map manipulation scripts allow for cleaning of the geometry of imported 3D maps. These can be used to remove static vehicles from a satellite-generated map, replace textures in the map with physically-based ones, and replace the satellite constructed foliage & vegetation with higher quality assets.

Contact Us

Questions or Comments?

For partnership and other enquiries, get in touch via our email contact@yorkshirevalidation.co.uk and we will get back to you shortly.

PlaceSim Autonomous Vehicle Simulator