Scaling Safe Autonomous Driving Software via Personalised Testing
Up to 2021, as players continue to race to define “autonomous safety” over $60 billion has been spent and tens of millions of miles driven across the automotive and tech industry. Meanwhile, on real roads 94% of crashes remain a result of human error leading to 1.25m annual road deaths around the world. What is missing in identifying the hardest challenges behind the safety of autonomous vehicles? Read on to get Rydesafely’s point of view and see how our solution of Personalised Testing can help.
[Note — we’ve kept this intentionally non-technical and easily readable to all. Please get in contact — tom@rydesafely.com — to go a level deeper on the challenge]
A few years ago, everyone said autonomous driving vehicles were ‘inevitable’. Yet here we are in 2021 and we are still not seeing them zipping around outside our houses.
Since the high profile crashes and accidents of 2018, the industry has redirected investment — around $20bn to be specific — to validate safety for their autonomous driving software systems.
However, no one has solved this expensive riddle yet.
What is holding back autonomous systems on roads?
In simple terms, Autonomous vehicles are made ‘smart’ by their underlying machine learning software and neural network models. Unfortunately these are black boxes. Without going too technical, this simply means that understanding what they’ve learned and where they’ll break is unsolved. This has massive safety and financial ramifications across the industry.
The industry has responded by blunt testing endless mileage in standardised scenarios in some sort of ‘expenditure safety race’. Waymo has contributed 20m miles on the road and a billion-plus dollars to 2021. Cruise also has two million miles on the road and every traditional automotive manufacturer is involved too to some expensive extent. Unfortunately, this blunt generic miles and dollars are not getting us any closer to truly scalable autonomous safety solutions.
The famous study below shows that we would need to drive billions — or even 100s of billions of miles — across decades for these current mileage test methods to prove autonomous system reliability. With 1.25 million lives lost each year because of human error, and $5–10B spent annually on autonomous development, we simply can’t afford this time or expenditure.
In short, generic and endless test mileage doesn’t work. We need personalised testing to the system at hand to get these systems on roads.
Moving Away from Endless Test Miles
But who can personalise safety testing?
Our engagements with policy partners suggest that individual companies are currently being left to define specific safety scenarios. Policy standards and other documents like ISO/PAS 21448, SOTIF and UL600 give guiding frameworks but — geography dependent — the industry generally takes the lead via public voluntary self-assessments (VSSAs) or historic reporting (California’s Disengagements below).
This industry-led approach partly makes sense from a privacy and technical perspective as Intellectual Property is critical for autonomous developers. As well as this, governments seem to have no desire to — nor have the technical capability to — write personalised tests specific to the neural networks delivered on roads.
But can our tech or automotive providers themselves safely define autonomous safety to their own systems?
There’s no doubt, individual industry providers are doing amazing work in autonomous development. However, they are failing in the last, hardest and narrowest work stream of ‘personalised’ safety to their specific system. This is because:
- They also have to deal with all the other development work streams from from the software engineering, Machine Learning Modelling (perception, prediction, trajectory planning to execution) to the sensors and even at times the vehicle hardware.
- Also, they only have access to their own silo-ed data or any simulated data they can think of to generate virtually and endlessly. They lack any sort of cross-industry view.
- They naturally hold some bias in their own definition of safety as an individual player with financial and reputational stake in autonomous delivery.
Analogies from Aviation around 3rd Party Safety Testing
In the airline industry, there is a generally accepted notion that an accident for one provider damages the reputation and bottom line of everyone. Therefore, if there is an airline accident — or high-risk potential for an accident — that risk is reviewed and traced between all providers via 3rd party agencies companies.
We believe that the same should apply to the automotive sector, and the industry apparently agrees with us.
Since the autonomous vehicle accidents in 2018, a number of new startups have entered the market with external safety validation software whilst still protecting automotive players’ intellectual property. The players in this space (full disclosure — our competitors) fall into three buckets.
In short, all of these players want to standardise safety for the entire autonomous automotive industry. At Rydesafely, we don’t think this is possible or the right way to get vehicles on the road.
The Importance of Personalising Testing Scenarios
The Uber accidents in 2018 and the Tesla Autopilot all had hugely different scenario parameters from the agent (a pedestrian pushing a bike vs. a sideways truck) to the use case (urban driving v.s. highway driving) to the time of day (night time vs. daytime), to the angle (right in front vs. walking from the right hand side).
We believe that the only way to solve for the most challenging and silent system failures is to personalise testing to the machine learning model at hand. This brings us onto Rydesafely.
Our platform is the only solution to auto generate personalized scenario-based testing to the specific neural networks that power the autonomous system.
Personalising safety challenges and scenarios to the specific autonomous system upends testing and moves the industry from endless quantities of mileage (costing billions of dollars over decades) to only testing the gaps in a specific system. This results in safety outcomes at 10x faster delivery and at 10% of the cost.
Rydesafely: Easily Deployed for Scalable Systems
So how does Rydesafely work. We’ll cover this in in more detail in our next blog post but in short:
- We deploy the Rydesafely app, docker-ized and on-premise so all data remains at the sole access of the client.
- Our platform takes in as inputs: i) The clients autonomous models and software they want testing ii) raw vehicle datasets — unannotated by humans — usually in the form of camera recordings but also Lidar / Radar sensors.
- Our first product then uses our novel IP to automate & personalise fault detection by comparing these 1000s of hours of raw driving data vs. the client’s autonomous software. This provides rapid fault detection results — sometimes called edge case detection — overnight and cluster the areas the software failed for easy review & fixing by an engineering team.
- That’s not all though. Once Rydesafely identifies the current vulnerabilities, we then generate entirely new test scenarios based off of these weaknesses. These scenarios create life-like scenes in simulation ensuring we capture potential risks that haven’t occurred yet in the real world. By hyper targeting only on the vulnerabilities, this avoids the generic scenario generation and endless mileage prescribed by our competitors for 10x the speed to deployment.
- Finally , we also create scenarios and sense check our work using our cross-industry proprietary data structure. This propriety asset can track the internal relationships between objects in the world to see where common accidents or failures occur. Simply put, it ensures that our client’s system are safe vs. global public and private data as well as ensure it aligns with government policy and standards. This pushes our automotive clients far beyond the silo-ed data they’ve collected from their own in-house vehicles.
How does this change the game?
This mantra of ‘Personalized Safety’ test approach will cut the time to market by 10x for automotive players. It also opens up additional billion-dollar markets for policy and insurers willing to underwrite risk based on more robust methods. This can move their industries from historic accident data only to instead writing policies for near misses and even failures that are imperceptible to the human eye.
We hope you now see how important a personalised approach to testing is. If you want to be at the forefront of autonomous deployment and save these multi-billion dollar industries then get in touch. We’re always excited to speak to new talent, clients and event potential investing partners.
Tom Stuart, Cofounder & CEO of Rydesafely has experience at the start-up Havn, Jaguar Land Rover and Samsung Electronics and loves to build transportation tech products. He and his team of innovators use personalised safety scenario testing to identify autonomous faults and then auto generate tests and scenarios to help get more life saving autonomous systems on the road. He can be reached at tom@Rydesafely.com.
