How can we stop autonomous cars from making bad decisions?

Source

Autonomous vehicles perceive their surroundings, navigate, and park — free from human interaction — and this technology is rapidly expanding. Recently, Alphabet Inc. (Google’s parent company) has demonstrated its autonomous vehicles to have driven over 5 million miles on public roads. However, very tangible concerns exist around safety and feasibility.

Case in point: Uber’s self-driving car resulting in the death of a female pedestrian in Tempe, Arizona in March 2018.

While engineers can program rules to determine a vehicle’s response to traffic situations, prediction of every possible situation is difficult, time-consuming, and not monetarily feasible. So this presents an interesting problem: how can we stop autonomous vehicles from making bad decisions?

Artificial Intelligence in Autonomous Vehicles

Autonomous vehicles come equipped with smart systems that generate tons of data. These systems include camera-based machine vision systems, radar-based detection units, driver condition evaluation, and sensor fusion engine control units (ECUs).

Furthermore, AI is expected to be the largest and fastest growing technology in the automotive market. However… there are cases where machine learning (ML) models are inadequate. (ML is a subset of AI!)

Solutions

1. Opt for built-in redundancy that is complimentary. So if one sensor fails, another sensor can still collect data and take over the system to make decisions.
2. We can design hybrid systems that encompass ML and non-ML components (Human input is powerful and will always be part of the picture!)
3. We can leverage expert domain knowledge where humans can input expert bias that constrains these systems to a few parameters.
4. We can also have expert-design algorithms using procedural.ai!

In fact, according to researcher Lex Friedman, predicting future behavior is a common and well-researched problem in AI.

PerceptionRNN component of Waymo’s ChauffeurNet predicts the trajectory of other cars. Here visualized in red is the past & in green is the predicted future.

Let’s look at this problem from the level of designing better training data for autonomous vehicles. procedural.ai stimulates imagined and real-world phenomena, allowing the user to account for landmarks, weather, traffic flow etc. from the perspective of an AI model. Our AI uses natural description to turn words into 3D scenes, which can be used as virtual test beds to test, train and validate ML models, ultimately helping autonomous vehicles from making bad decisions.

References:

Thanks to Drago Anguelov for the solutions expressed in this article. You can watch his talk here.

Read our other articles:

Is artificial intelligence a threat to humanity?

Can computers think like humans?

Will AI kill all the jobs?