On 14 March, 2018 we had one of the awesome data science events organized by the institute of computer science. This time topic was about autonomous vehicles, a hot and growing field which benefits AI more and more. The most interesting of all were the Andrej Karpathy’s talk from Tesla, and Lindsay Roberts from Starship.

Quick Overview

• ISEAUTO (TTÜ ja SILBERAUTO AS): self-driving car project of the Estonia - it looked like more as an autonomous minibus, AI-powered using ROS and its modules, mostly applied by students as their academic projects!
• Starship presented by Lindsay Roberts - an interesting project for package delivery using autonomous robots. The talk focused on localization of the robots mainly by employing intertial sensors and landmarks captured from edges in the image. Bottleneck of the system (huge amount of the features detected and used to this date). The main algorithm under the hood for this is a particle filter!
• Milrem Robotics: Modular Military Vehicles (structure), Autonomous exploration and navigation in Estonia’s forests.
• A presentation on legal matters and need for adapting new regulations to the new technologies.

Photo by Dima Fishman

• And finally Andrej Kaparthy’s talk that I put it into an agenda:
• Tesla autnomous cars are the largest group of robots employed in the world
• Tesla, toward a sustainable, autonomous, and safe transporation
• Tesla cars are giant animals of tiny brain mapping sensory inputs to the steering and acceleration
• A progression towards software 2.0 from 1990 – 2010
• software 2.0 a concept that Andrej has written on recently
• Massaging the data instead of massaging the code (yet another software 2.0 implication)
• Conundrums towards solving the autonomous car problem: rising philosophical questions
• Not just dealing with rare classes, but also rare scenarios (very odd incidents, few examples to learn from)
• Idea of autonomous vehicle started in 1989 Pomerleau Alvinn; an autonomous land vehicle
• Many machine vision practices are obsolete now, i.e. SIFT, however SLAM/visual-odometry, and many other techniques will be around for now
• How debug software 2.0: massaging the data! interpretable stack
• Future progress of Autnomous vehicles : development of the infrastructure (i.e. induction lanes) + smarter cars
• My car is always charged, plugged on both ends (unique experience, compared with the era of going to gas station!)
• graduated from numpy, using PyTorch now!

My remarks

• From hardcoded-automation to datacoded-automation (our new journey towards automation)
• It is interesting that our technology evolves like a huge neural network. It has got multiple passes until it gets to work (its optimal point). Many ideas from 70s or 80s even before emerged, however took many year and now (after a couple of passes - like backward propagation Lol) it has gotten to its golden time applied and used many places – neural networks are subject to this too
• Starship is a very good example of automation as a real-world example. It demonstrates how capable we are in full automation in a real world scenario.
• Perhaps we to this day we are capable of 30% automation in our life and industry. Years to come are going to flourish the automation era
• In 100 yrs time automation is no longer of our concern

More pictures of the talks on I+EE facebook page!