Turing Motors Approach: Object Identification
This article is part 2 of 8 in a series written by Turing Motor Company’s CEO: Ned Goodhue. This series outlines the Turing Motors approach to bringing level 5 autonomous vehicles to market.
Identifying Objects is completely different in an AI environment.
AI is all about the real world. Our team uses pure AI plus common sense to teach our cars how to drive.
In July 2017, Volvo (then a leading AV team) went to Australia to improve their big animal model (1).
The team “admitted” that their “self-driving” cars are confused by kangaroos.
This is the way many AV teams get confused. That approach is difficult to implement and to understand its results.
There is an “AI in theory” and an “AI in the real world”. That is the way many teams have given up and went back to simulation. Here is why:
- Humans have the unique ability to know a specific object very accurately. A 6-month-old baby will know its mother out of a group of people who look like its mother. But humans are not able to keep many objects in their brains at the same time. Similarly, when driving, humans can only look in one direction at a given time, usually forward. High Accuracy — Few remembered
- Computers — AI is not very good at identifying a specific object but can keep hundreds of objects in their data set. Conversely, cars can have sensors and cameras all around the vehicle and simultaneously be tracking every direction. Many Objects — Hard to classify.
Turing Motors has a simple solution — we increase our knowledge about a specific object by its behavior over time.
In the Turing Motors world, we meld both methods to understand many objects. It is amazingly simple.
- Computers can re-calculate fast. Our cars will recalculate an object’s trajectory, its intent, 3 times per second.
- AI can see the objects’ movements to improve its knowledge of that object.
- Object + Motion + color = knowledge about the object can create incredibly accurate identification of the ‘object’ without giving up speed or the number of objectives identified.
The calculations.
This is what our cars know in:
.07 sec Is there an object there?
.13 sec If “yes,” what is the direction of its movement and its speed?
.20 sec Is it accelerating? Is it turning?
.27 sec Is it accelerating more or less? Is it turning more aggressively or less?
What have we learned?
Is it moving?
If not, what is its intention?
In 1/3 of a second, we will have created a trajectory map, like those used to track hurricanes, with probabilities, so the car can decide the safest maneuver given the object's trajectory.
Using both forms of AI, we can identify a substantial number of objects with the accuracy of a person and the speed of a computer.
Next time
In the next article of the series, we will discuss the role of Simulation, Inference, and Making good decisions in Level 5 AVs.
Our technology is vastly different from any team that we know of.
It is 3 times the power at 1/3 of the cost.
This article was written by Ned Goodhue. Ned is the Founder and CEO of Turing Motor Company. Ned is passionate about all things AI. Ned is also passionate about aviation, travelling, understanding different cultures, theatre and reading.
Feel free to leave a comment, whether you agree or disagree, we encourage a collaborative discourse. Thank you.
-Turing Motor Company
References
- Naaman Zhou. (2017, July). Volvo admits its self-driving cars are confused by kangaroos. Retrieved June 13, 2021, from the Guardian website: https://www.theguardian.com/technology/2017/jul/01/volvo-admits-its-self-driving-cars-are-confused-by-kangaroos
