Intel VP on teaching cars to drive at IoT World 2017: it's just like teaching your own kid

Home > News > Content

【Summary】At the Internet of Things World 2017, Intel VP and General Manager Jeff McVeigh has made a vivid analogy of how to teach cars to drive by themselves.

Original Claire    May 19, 2017 1:44 PM PT
Intel VP on teaching cars to drive at IoT World 2017: it's just like teaching your own kid

At the Internet of Things 2017, the World's largest IoT event held in the heart of Silicon Valley, Intel VP and General Manager Jeff McVeigh has made a vivid analogy of how to teach cars to drive by themselves: "It's just like teaching your own kid." Futurecar reporters were on site to cover this important keynote.


Teaching a teeanger to drive in a lifecyle

Using his 16-year-old daughter as an example, McVeigh listed four processes that a human being uses to learn how to drive: gain experience under parental guidance or driving school education, practice with over 100 hours of driving behind wheel, test and launch by taking road test, localize and retrain to get adjusted to new conditions and new rules of the road.

A human being is using perception, reasoning, and prediction in drive-learning process, but what about cars?

"Suppose, the driver is the software that's running the car. The question is, how do we teach the software to do the right thing, to behave properly? To deal with the real-world dynamic challenging environment that we need to learn with our own lives?" He asked.

What an autonomous vehicle should have

Radar, cameras, LiDAR and ultrasound are providing the car's perception to sense and detect the surrounding world; vehicle-to-vehicle(V2V/V2X) communication and high-definition maps are modeling the environment so the car can recognize it, then the car is fed the instructions, or learns by itself to develop its own driving strategy, leading to motion control.

Like a teenager who never drove a car before, a self-driving vehicle needs to accumulate over 1 million miles of real-world sensor data to "know the world". For example,  what does a tree look like, that object should be labelled as a stop sign, there's a house but not a person in the front, etc. It's a process of getting familiar with varied road conditions and locations, and feeded with labelled and segmented ground truth.

In the second phase, the car is practicing driving by "algorithm partitioning, artificial intelligence/deep learning, 10X the computing requirements of L1/L2 autonomy", simply put, with the humongous amount of data collected, the computer is trained to figure out on its own (machine learning) what an object or environment is after processing the data in its own logical system.

Then the car needs to carry on a closed-loop simulation, functional safety certification, human-machine interaction. The third step, just like unexpected driving conditions, can happen at any time, in any way. The car needs to learn anomaly detection in a changing environment.

"We'll mix the real-world data gathered by the fleet vehicles with simulated data. That simulated data is important because even driving a million miles you might never encounter a rare scenario….it's important to utilize simulation to fill some holes to ensure that you're achieving the right performance." McVeigh said.

Making a perfect autonomous car doesn't only mean it's machine perfect, interaction with humans is also key in building up connections.

"It's not just about making it safe, but ensuring that passengers understand what the car is doing and build up confidence to trust the vehicle."McVeigh stressed.

The localization and retrain process

Often times, changes occur. For example, some new construction, a new road, or something else that needs to be updated in the mapping system. The information need to be fed back to the data center.

McVeigh calls this phase the retrain process. He added that this also applies to driving in a foreign country where transportation signs, traffic rules, and conditions can be very different.

"Driving decisions that happen in different environments need to be tailored in hours of training and teaching process."

Why connected car is important?

Cars can share their experience in the future world. An autonomous car's computer system will not only be fed data by its engineers, but also data from other cars on road. Their experience with the environment can also help an individual car "see" around corners, learn from others, build up the network and enrich the data center resources.

"Just like my 18-year-old daughter would teach her 16-year-old sister the do's and dont's in driving, the shared-experience will make her avoid bad driving habits. And this is collective intelligence."

Prev                  Next
Writer's other posts
    Related Content