A few weeks ago, Indian autonomous driving company Swaayatt Robots claimed to have achieved Level 5 autonomy. The company announced this big breakthrough in a video, showcasing autonomous driving of the car at night at a toll plaza amid complex traffic dynamics. 

The video, however, failed to convince the sceptics as here was a company claiming to have successfully developed Level 5 autonomy at a time when Tesla, General Motors and other bigwigs were still languishing at Level 2. 

To clear the air, AIM did an exclusive interview with Sanjeev Sharma, founder and chief of Swaayatt Robots, who confirmed, “If we look at the demo conducted at the toll plaza, the vehicle showed Level 5 capabilities.” 

“At night, there were no traffic rules. You could see trucks going zigzag, overtaking our vehicle randomly. Also, there were trucks parked here and there. Our vehicle, arriving at the intersection, slows down for the speed breaker and then decides which lane or gate to commit to,” he said, noting that this capability falls under level 5 autonomy.

Moreover, Sharma said their technology is capable of negotiating bi-directional traffic on a single lane. “No other company deals with this; they just come to a complete halt. Our vehicle just took a small detour even when the object was coming from the wrong end. Now, we are able to do this because this intelligence is inherently embedded.”

As for the level game, he said that he leaves it up to the people to decide if this was level four or five. In fact he said that he would be fine even if they wanted to mark it level three. “However, wherever we felt that a claim had to be made, we made the claim, specifically in toll plaza navigation, asserting that this is level 5.” 

Tech Behind This 

Sharma said that Swaayatt Robots is heavily invested in reinforcement learning and inverse reinforcement learning. He considers Wavye AI to be their toughest competitor, given that Wavye AI also employs reinforcement learning. Moreover, he said that Swaayatt Robot’s current focus is to get rid of as many algorithms as possible. 

According to Sharma, while other companies are working on developing algorithms for obstacle detection and the intent of the other vehicles, they are working on eliminating the obstacle detection algorithm in autonomous vehicles. Their focus is on creating a decision-making system with an inherent understanding of the world’s context, without the need for explicit computation to detect various obstacles or road intent. 

Recently, a Bengaluru developer converted his modified Maruti Alto K10 into an autonomous vehicle using a second-hand Redmi Note 9 Pro and an open-source fork of Comma.ai’s OpenPilot.

Regarding Comma.ai, Sharma mentioned their use of behavioural cloning. “Behavioural cloning means you try to mimic or clone the human demonstrators or experts. This is the approach taken by Comma.ai and some other startups working on this model of autonomous driving, but it’s not scalable – it’s a dead end already,” he said.

Speaking about Tesla, he said that they use autoregressive reinforcement learning, which could be scalable, but a bit doubtful as training and running ARL models can be computationally expensive, requiring powerful hardware and significant resources. 

Sharma said that Swaayatt Robots is trying to challenge the three major pipelines of autonomous driving in the “classical sense”, which are perception, localisation and mapping and planning.

“Since 2019, with the advent of Wavye AI, people have been discussing autonomy without relying on maps. Localisation against maps necessitates the use of maps, but we achieved everything without them. We were the first technology to enable vehicles without the reliance on high-definition maps. In 2017, we implemented multi-RL agents without requiring any maps,” he said. 

Furthermore, he added that they have autonomous vehicles with mounted cameras; hence, they don’t require external data supply. They have pipelines for automatic data labelling and use generative AI to create specific scenarios, though not on the scale of NVIDIA. Their setup includes 8 side-looking cameras, 2 lidar units, and 2 additional cameras on the bumper. We plan to add 4 more cameras to the rear. 

What’s Next? 

Sharma revealed that the team is actively working on an undisclosed ‘X’ paradigm, intending to showcase how autonomous vehicles will independently acquire various skills. This demonstration is scheduled for the month of February.

He said that the autonomous driving tech market is projected to be a trillion-dollar industry by 2030, with only a maximum of five companies expected to survive. “Our goal is to secure 25% of the market by 2030 and evolve into a multinational corporation,” he added. 

“I believe that in a year or so, we will have a pan-India presence and will have raised $1 billion. Currently, we have only $3 million in funding, so even if we achieve sophistication in a specific algorithmic framework for a certain problem, it can only be demonstrated on a limited scale,” he concluded.