Swaayatt Robots: Pioneering Reinforcement Learning in Autonomous Driving

Bhopal-based Swaayatt Robots isn’t your traditional autonomous mobility company. The startup focuses on developing self-driving technology for unstructured environment conditions and India’s road network is full of such environments.  

In the thick of it is founder and CEO Sanjeev Sharma, whose interest in the field of robotics was born way back in 2009, when he watched the videos of Team MIT at the 2007 DARPA Urban Challenge. With time, he knew that he wanted to hone in on research to enable autonomous driving in the most difficult traffic environmental scenarios, but it wasn’t until 2014, when Sharma deferred his PhD at the University of Massachusetts for a year, that he established Swaayatt Robots.

Fast forward eight years and, despite knowing much more about autonomous mobility than in 2014, safety continues to be a huge challenge. Even before we think of the purchasing and operational cost, we’re quite some time away from solving for driver safety in an uncontrolled and unstructured environment — but Swaayatt Robots is trying to fix that. 70% of the research done by the company is focused on motion planning and decision-making under uncertainty, with the goal of developing algorithms that can work in the world’s most difficult traffic and environmental scenarios.

Swaayatt Robots focuses on two main aspects — traffic laws and traffic dynamics. In India, for example, traffic dynamics are much more complex than in most developed countries around the world.

“There is a lot of randomness in the behaviour of traffic jams, the structure of the roads and environments in India,” says Sharma.

“If we could develop the algorithms that can ensure the safety of autonomous vehicles on Indian roads, where the traffic dynamics are so complex, then those algorithms would be able to guarantee the safety of autonomous vehicles anywhere else in the world as well.”

Paving the way for Safer, More Cost-Effective Autonomous Technology

“We were heavily focused on researching in the areas of motion planning and decision making under uncertainty,” explains Sharma.

“Along with that, there was also the problem of cost and scalability. Let’s face it, LiDARs are very expensive. We started to ask ourselves one question — how do we ensure that the autonomous vehicles can perceive their environment and that too complex and random environments like in India, while still being cost-effective? That’s when I started researching perception and began to develop algorithms that can help autonomous vehicles semantically and contextually understand their environment using only off-the-shelf cameras.

“These two were the primary objectives when starting Swaayatt Robots. One was safety, where we were researching motion planning and decision-making under uncertainty. The second was the area of perception, where we were developing algorithms to enable autonomous vehicles to perceive only using cameras. The third idea, where we dedicate 5% of our research capabilities, is to enable autonomous vehicles to navigate in purely unknown and unseen environments, thus getting rid of the high-fidelity maps required for autonomous driving. There are only 3 or 4 companies around the globe that currently do this, but we have two Proof of Concepts (PoCs) already.”

Swaayatt Robots is starkly different from companies in the Indian autonomous mobility space.

For starters, the development of algorithms is an R&D challenge, one that the company views as such. It develops its own algorithms, while most other companies tend to rely on state-of-the-art algorithms developed by research labs globally. As hundreds of papers on this topic are published each year, most players tend to prune these algorithms, reduce the computational requirements, and fine-tune them to solve their purpose.

The result is that Swaayatt Robots has developed algorithms that are at least 10-20 times more computationally efficient than the state-of-the-art, 10-14 times less expensive for some tasks and with a reduced computational burden by 10-40 factors, significantly reducing the energy requirements and carbon footprint as well.

“So far, we have developed 47 different novel algorithmic frameworks, of which the most prominent is the Multi-Agent Intent Analysis and Negotiation Framework,” says Sharma.

“What it does is it allows autonomous vehicles to know what other vehicles on the roads are trying to do and take an action taking the future actions of the other vehicles into account. Many other companies are trying to solve this, but it has one major problem — it is computationally very expensive. Typically, the computational requirement grows exponentially with the number of obstacles or the number of moving obstacles or agents that you have around the vehicle. Now, I’ve been researching in several fields of theoretical computer science and applied mathematics, including the field of topology. We’ve been successful at developing an algorithmic architecture that has been able to reduce this cost from exponential to quadratic for some cases, which happens to be a major breakthrough in the field of multi-agent systems and multi-agent reinforcement learning. We’re now trying to further bring it to super-linear for some special configurations.”

A Long Road Ahead but Light at the end of the Tunnel

While Swaayatt Robots’ accomplishments sound impressive, the company has been on a long and winding road.

For starters, the company has had its share of challenges while seeking funding. As Swaayatt Robots’ approach is R&D-intensive and doesn’t involve relying on pre-existing algorithms, the company was a new proposition for investors, most of whom had a hard time coming to grips with this. But the team’s patience and perseverance paid off.

“In July 2021, we received $3 million at $75 million post-money valuation from a North American investor. They have invested knowing that this is a single-founder company, one with just 7 members in size at the time of investing, where the approach is R&D-intensive,” says Sharma.

And that isn’t the only challenge.

“Getting the right research talent in India is a challenge,” Sharma explains, adding that the present-day education system is to blame for this. The problem statement is such a powerful one, that Sharma created a second start-up, Deep Eigen, to solve it.

“Deep Eigen exists because we want to revolutionise the education system in India. Getting developers in India is not hard at all, but when it comes to getting researchers, we have to hire people who have done a Masters in Science (MS) or a PhD from universities abroad.

“This is a big challenge because, in our field, R&D is at the heart of it all. We have to be very careful about the people we hire and check if they have an understanding of the market or not. That’s why our hiring is very slow. When it comes to programming, there are awesome developers you can hire, but when it comes to the fundamental development of a new mathematical idea, that’s where the challenge begins. Our plan was to have a team of 17 after the $3 million funding round, but we’re just a team of 12 still. That said, we’re slowly and steadily seeing this problem resolved. We’ve had interest from people who are doing their MS from outside India, who’re coming to join the company.”

If DeepMind and Mobileye were pioneers in the field of reinforcement learning, Swaayatt Robots is probably the third company to follow. Sharma is confident that the company is here to stay and it won’t rest until it brings autonomous cars to Indian roads, in a safer, scalable, and cost-effective way, whether it is through technological breakthroughs, or being first in line to collaborate with the government on forming relevant policies.

Speaking about what we can expect to see from Swaayatt Robots in the near future, Sharma adds:

“There is a RACER programme being conducted by the DARPA (Defence Advanced Research Project Agency) in the US. The goal of this programme is to develop algorithms, which can enable autonomous vehicles to navigate at human-driven speeds off the roads. We already have a proof of concept of this ability existing since 2014, and my research was in the area of unknown and unseen environments. Over the next few months, our goal is to prove that we are able to enable autonomous vehicles to navigate at human-driven speeds off the roads using a minimal amount of sensors, which directly challenges state-of-the-art and whoever is participating in the RACER programme.

“Our six-month goal from today is to make autonomous driving Level 3 prevalent on Indian roads in any city. Eight to nine months down, we want to ensure that we are able to execute 100km/hr autonomous driving on Indian roads. We hope that when we demo this, we would be the only company that has the capability to perform such an experiment. Our experiment would be performed on mountainous terrain, as well as have 120-degree turns. We want to be able to prove multiple factors – that we have algorithms that can enable autonomous driving without high fidelity maps, that allow autonomous vehicles to perceive the unstructured environments, and finally, that can ensure safety even in such road scenarios. These three points are crucial in proving why Swaayatt Robots was established and why it is a thought leader when it comes to autonomous driving.

“Moving forward, our funding requirement is around $30-40 million. Our Series A target is around $127 million. Once we have achieved this technological feat and showed 100km/hr autonomous driving on Indian roads, we don’t see getting $127 million to be a problem for us at all.”