Sanjeev Sharma

Enabling autonomous vehicles perceive their environment using only off-the-shelf cameras has been a long term research objective at Swaayatt Robots (स्वायत्त रोबोट्स).

This demo highlights the capabilities of our on-road perception system which is able to detect obstacles, road boundaries, lane markers in images, as well as compute depth of the complex scenes in the environment. The output shown in this video is end-to-end raw output from our deep learning system, without any post processing.

The current system, with joint computation of obstacles, lane/road boundaries, and depth, works at 30 FPS on an embedded GPU in our autonomous vehicle, and can achieve higher FPS with further optimization -- which is currently a research in progress.

This system is being scaled up for both the day and night operations, and we will showcase its strength towards enabling autonomous driving on a mountainous environment with unpaved roads, in the absence of any delimiters.

#deeplearning #autonomousdriving #autonomousvehicles #machinelearning

Ability to negotiate tight-obstacles on any kind of roads is one of the key abilities that autonomous vehicles should demonstrate to be able to be scaled sustainably, and to achieve safety without compromising on the energy consumption.

On September 21st we did an off-roads #autonomousdriving demo “Driving where no Autonomous Vehicle has driven before!“ where our vehicle traversed the kind of terrain in a manner that has never been done before by any autonomous driving startup.

On November 15, 2017 we did a very tight dynamic obstacles avoidance demo with the environments and obstacles being adversarial-stochastic-complex-unstructured in nature, using our multi #reinforcementlearning agents based framework that I was working on that that time -- which was world's first demo of end-to-end deep reinforcement learning applied to real-world #autonomousvehicles, that too for such kind of navigation tasks.

Over here in this demo, again one of the laps from Sep 21st demo, we show a custom designed motion planning and decision making algorithmic framework, that is able to negotiate tight passages, without stopping the vehicle completely. This work is further being scaled with unsupervised learning.

At around 28 seconds the vehicle encounters a situation where a car was parked on the left, and a biker was approaching from the other end on the right, leaving not enough gap for the vehicle to pass with a safe enough margin. Usually in such situations, typically autonomous vehicles in the West, would come to a complete halt on the legal driving side, and wait for the other obstacles to pass -- this can also create a dead-lock situation, like what is happening in SF with Cruise, Zoox and Waymo autonomous vehicles, whose software is not properly designed and makes lot of assumptions. Often, the planning software in such vehicles designed by those companies is running at such a low frequency (<10 Hz) that is just not safe at all. This gives rise to another problem to solve -- like predicting the future of other obstacles, etc, which again is a dead-end.

Earlier this year Crusie showcased avoiding dense traffic in SF, where it could be easily seen stopping for every possible obstacle imaginable.

Here, in our demo, it can be seen that the vehicle never comes to a complete halt and tries to, and successfully does, find a passage through the obstacles, on a single lane road.

We are scaling this algorithmic framework via inverse unsupervised learning, and very soon its performance will match that of our Nov 15, 2017 demo. The difference being that this framework will be a non-holistic decision making framework, whereas the former was a holistic (every control command was computed by multi RL agents end-to-end). Around January we will showcase this framework converging to a near holistic framework , i.e., when activated by the vehicles' probabilistic decision making framework, it will compute actions without taking input from perception software.