Autonomous driving through very dense dynamic traffic, with extremely tight-complex-stochastic traffic-dynamics on sub-urban roads, connecting to an open ground, with absolutely zero traffic-rules. This is the most heavily cluttered environment where we have tested our #autonomousdriving technology, presenting many of the adversarial negotiation scenarios as well, throughout the autonomous... navigation task. This demo was done at the Mata Baglamukhi Madir campus in the city of Nalkheda, in MP, India, and was done in the presence of heavy police forces deployed that day on the ground, as can be seen in our demo. Our autonomous vehicle starts from the temple with a generic open environment, with zero traffic rules, with very narrow corridors created out of barricades for vehicles movement by the security forces. In the corridor no two vehicles can pass through at the same time, and our vehicle was tasked with driving through this corridor, while negotiating its way from any traffic, two-wheelers, or pedestrians it faces, with dense presence of bikes and cars on either side, presenting a very challenging environment for #autonomousvehicles. The vehicle exits the open area, and then assumes generic dual lane navigation, avoiding both static and dynamic obstacles, before encountering a police check-post, where the vehicle is supposed to wait if the barricade is closed, and proceed if open. Upon exiting the checkpost, the vehicle negotiates a traffic-intersection with stochastic and adversarial driving behaviour of other vehicles on the road. Our vehicle continuously faced heavily cluttered traffic scene, where entities on the road can execute a random driving pattern, making the decision making task very challenging. We did the demo over a period of two days, successfully executing multiple (30+) trials in this setting. This demo was again a culmination of our prior works and demos: Kankali Kali Mata demo, on-roads, bidirectional negotiation capability on single lane roads, and open environment Level-5 negotiation capability as showcased in our Toll-Plaza demo. We again scaled up classical decision making and motion planning algorithmic framework, to adapt to such a level of density of obstacles on the road. This framework is further being scaled up with #reinforcementlearning and unsupervised #deeplearning at Swaayatt Robots. We will again do a demo in the month of June here, showcasing autonomously acquired skills to pave the way for Level-5 autonomous driving, and to solve the Level-4 autonomy problem by the end of 2024. #MachineLearningshow more

The sad part is that we've become so numb to our horrendous infrastructure... That we've created an AI innovation to navigate through this mess!

My research statement since my undergrad days has been to solve the problem of autonomy (which is the toughest AI problem of this decade, and was the toughtest problem back then as well) in absolute chaos. While this may be horrendous, as a researcher, that is somehow, a boon for us, as of now, so long it lasts :)

little confusion. agr kisi ko thok di. to pitega kon ?

I think the CEO of the startup :(

This is amazing work. It is evident that you are building this from scratch.(unless you had a dwarf in the car) The part where the scooter guy weaved in front of the car got me thinking how the car would react to a surprise doggy. Would love to see you work towards that. Will save kids lives also. This video will sell it out of the ball park because it builds trust through providing safety. Amazing work again! Keep doing what you do!

Thank you so much Vishal!! 🙏 Yes, navigating through the corridor was a challenge, and this is the narrowest road (if we can call it a road) we have tested this motion planning and decision making framework. Over the past few years, since undergrad days, my research focused on enabling autonomous navigation (primarily, then, focusing on motion planning and decision making under uncertainty, from mathematical optmization, RL, and ML perspective) in previously unknown and unseen environments. One key motion planning and decision making frameworks, that uses two RL agents, which has been demoed in our campus demos (here -- which is currently not a part of this demo done over here, works at nearly 2000 Hz on a single thread of core-i7 cpu (on really old ones) and can react very quickly, given we are solving the autonomous driving problem in a classical sense. This algorithm was rigorously tested throughout 2022 as well where we killed the perception horizon of the vehicle and forced it to reach to sudden appearing obstacles in our campus. Currently we are also extending that work via deep RL, and building an abstract learned representation where that RL work, and this current motion planning and decision making framework that enables very narrow lane negotiations, can be integrated -- to solve the kind of problems you have mentioned. There is a lot going on in the background, and we are actively raising another round to scale up all of this work, and to extend our previous work, and integrate all these components as well. One of our current research also focuses on killing the requirement of explicit perception algorithm altogether, and we are working on this front off-roads. A lot more to come in the coming weeks and months.

Can it read facial expressions? That is the final step required for Indian roads, when seeing a pedestrian or driver who tells you which way he's going to go, or if he's going to cross, through facial expressions.

amazing 🤩 congratulations! during testing, what failsafe do you have to intervene in case of likely collision?

Thank you so much! Notice, there's always a safety driver. In very chaotic traffic, usually there is person on the back seat as well, alerting of the situation behind the vehicle so that if anything goes wrong, the safety driver can intervene. Here -- you can look into our January 2023 demo, when we're doing off-roads for the first time, and that too at night in very dense fog. At that point, single lane inflow traffic negotiation off-roads was not a solved problem, and was under R&D, and thus during the demo you can hear team taking about being ready to intervene and take control of the steering as there was a bike coming in from the other end. So before we test anything in real-world, it has already gone rigorous testing.

Whoever solves autonomous driving in India, will rule the world.