Continuing on my earlier posts on a few interesting data sets on some select startups (and scale ups) in the autonomous driving eco-system………….
An end to end deep learning structured autonomous system – now that is a great holistic way to build up one for mobility. While most of the players are applying AI with a piece meal approach, mostly only for perception, Drive.ai is building up its AI software for autonomous vehicles via a holistic approach structured entirely around deep learning. In fact it sees deep learning as the only viable way towards autonomous driving.
Set up in 2015, the founders are deep learning experts from Stanford’s AI lab. Besides perception, there are other areas too which exhibit patterns and are apt for deep learning leverage – decision making, motion planning for example. However one of the major concerns of deep learning systems is that they are kind of black boxes – we do not intuitively know or understand the route/reason to that decision spurted out by the deep learning system. This would have been perfect fit for a rules based world but that’s not the realm we live in. Hence this is an interesting challenge that Drive.ai is looking into.
Data annotation i.e. the labelling of boxes around various objects (presented in the humungous amount of data sets generated) that could be relevant to the deep learning algorithms is presently quite an intensive exercise – with humans involved. Drive.ai is using deep learning to enhance automation for this.
An aspect that piqued a special interest for me is that this is one company which is factoring in the social side of driving too into the autonomous driving equation, especially with regards to other road users. It wants its vehicles to replicate the communicative side of human driving experience too e.g. gesturing people to cross, head nods etc.!
Below is the Drive.ai snippet from my brief:
Look out for my 3rd of 5 parts post on this topic later this week.
