Billion Dollar Question: How good is the perception system in your AV?
What the Current Metrics Tell Us
In our last post, we shared how BlueSpace is accelerating autonomy with our motion first approach to autonomous perception. Our 4D Predictive Perception’s ability to measure the true motion state of all objects in all environments provides the AV/ADAS stack a level of motion accuracy previously not possible through traditional deep learning approaches. Without reliance on prior training or HD map, our technology is truly a game changer for achieving capital and time efficiency while scaling to new geographies and vehicle form factors. Yet, to effectively scale, AV/ADAS players must be certain, and prove to other public and regulatory stakeholders, that their systems are safe. BlueSpace provides this certainty through our focus on industry-leading motion accuracy.
The automotive industry still hasn't come to a consensus on what metrics are appropriate to track and evaluate the safety and performance of an AV system. Some publicly available reports only point to superficial metrics including:
# of real miles driven
# of miles/hours driven in simulated environments
# of disengagements
Everyone in the industry knows not all miles are equal (highway vs. urban city streets), and disengagements reports can easily be influenced by the ODD and other factors. We need to dig deeper. While the metrics above provide reference points, they don’t get to the heart of the matter - how accurately is the AV system perceiving and predicting the surrounding world? The AV industry must provide all stakeholders a higher level of trust and comfort that meaningful metrics, 3rd party validation, and accurate measurement of error bounds were vetted before we put our loved ones in the vehicle. It’s not the total miles driven, but the mile that you will be driven in that matters to you.
Industry-leading Motion Accuracy Performance
Accuracy is typically defined as, “the degree to which the result of a measurement conforms to the correct value or a standard.” Accuracy is imperative in the automotive industry, where a few milliseconds can be the difference between life and death.
We tested our software on motion accuracy, using real-time kinematic positioning (RTK) equipment that offers a 2cm/s accuracy level. Our team then compared our performance metrics against submissions to leaderboards released by Nuro and Argo. The result: Our 4D Predictive Perception provides motion accuracy 10x-100x higher than industry benchmarks.
In addition, our software provides tight bounds on future predictions that are based on physics and mathematically derived algorithms, rather than inferences derived from a training based approach. This allows motion control and planning teams to plan around tangible error margins.
Downstream Impact of Highly Precise Perception & Prediction
The ultimate goal of autonomy is safe and efficient movement in a dynamic environment. First principles reasoning suggests this goal is best achieved when the motion planning is based on the most precise measurement of the current and future states of surrounding obstacles.
That is why at BlueSpace, we care not only about false positives/negatives (“Did you see an object or not?”) but more importantly, motion accuracy (“How well did you see the object?”). With BlueSpace’s highly accurate motion measurements and predictions for velocity, acceleration, and yaw rates, motion control systems can make decisions with higher precision and with lower latency than ever before.
Unlocking More Robust and Reliable ADAS and AD Features
While sensor configuration/types may be different across different manufacturers and form factors, the software algorithm that processes the sensor data will play a pivotal role in determining the safety performance and capabilities of the AV system.
Our next post will share examples of how the motion accuracy and time advantage provided by BlueSpace’s Predictive Perception software unlocks new ADAS and AD applications. Stay tuned to learn about the “Two Second Advantage.”
Can’t wait? Reach out to our team at partners@bluespace.ai to demo our technology and see for yourself.