Swaayatt Robots developing autonomous driving tech in India

Bhopal-based Swaayatt Robots has been developing and demonstrating their autonomous driving technology for 7 years – including demos for lane detection, night driving, bidirectional traffic negotiation on a single lane road, off-road driving, and toll-plaza negotiation. The company aims to showcase an end-to-end negotiation of the daytime traffic in the upcoming months.

In 2021, Swaayatt received a seed investment of USD 3M at a valuation of USD 75M. Founder Sanjeev Sharma believes that 4 to 5 autonomous driving technology companies will survive globally by 2030, and Swaayatt aims to be among those select few.

Would achieving level 5 autonomy, proven and validated in various markets, result in safer road conditions?

For sure. Improved safety is a natural byproduct of autonomous driving technology. The ultimate goal is to achieve Level 5 autonomy, which means developing self-driving vehicles capable of making decisions. This will significantly reduce errors unless there are faults in their sensors. This advancement holds great promise, particularly in addressing the prevalent issue of driver errors leading to accidents, as observed in countries like India, where drunk driving is a major cause of truck accidents. Implementation of autonomous technology can vastly improve road safety globally.

Moreover, beyond civilian applications, these advancements hold immense potential in the defence sector, where they can save lives by undertaking complex and hazardous tasks. Additionally, application in search and rescue robotics underscores the diverse benefits of autonomous technology. Considering the broader spectrum, applications like drones further highlight the versatility of these robotic systems, promising safer and more efficient operations across various domains, including civilian transportation, defence, and emergency response.

Please help us understand the current readiness level of your technology.

Currently, we can handle almost every scenario (as demonstrated in parts) except for day time driving in city traffic. A holistic demo from A to B requires significantly more resources and funding. Thus, we showcase individual demos such as toll plaza navigation and off-road driving.

Our demos include bi-directional negotiation on single-lane roads and complex toll plaza scenarios, showcasing the vehicle’s ability to make optimal decisions sequentially. Additionally, our Off-road research aims to achieve high-speed autonomous driving without relying on high-definition maps or explicit perception algorithms. We seek to embed intelligence in the decision-making layer to enable end-to-end decision-making without explicit environment perception algorithms.

These advancements address major challenges in autonomous driving, demonstrated within our limited resources and funding. While companies like Waymo and Tesla offer solutions for specific segments, we aim to solve the entire problem of autonomous navigation across different verticals.

Please tell us about India’s regulatory framework related to autonomous driving.

The Motor Vehicle Act hasn’t been updated since 1988, lacking any mention of autonomous driving. Major companies are integrating autonomous systems into their vehicles, like the XUV 700, without specific regulations. In case of accidents during testing, the responsibility lies with the driver, similar to conventional driving.

We inform the local authorities about our demos, a safety driver is required for legal driving. There’s no current law prohibiting autonomous driving or specific mention of it in the Motor Vehicle Act.

You are developing algorithms to enable the use of off-the-shelf cameras without using RADARs and LiDARs. Please help us understand your approach.

Our previous demos, such as one in November 2017, showcased our ability to enable end-to-end navigation using only two front-facing cameras. Over the years, we have developed algorithms that perceive environments, both day and night, without a LiDAR. For instance, our lane detection and generation algorithm has been demonstrated in over 10 demos since 2017.

Initially, purchasing LiDAR was expensive, driving us to develop algorithms to extract contextual information via visual data from cameras without relying on LiDAR or maps, which impact operational costs. We have heavily researched decision-making and motion planning to handle chaotic environments, integrating reinforcement learning to make LiDAR and RADARs redundant for autonomous driving.

Our off-road demos typically use LiDAR. The use of LiDAR for on-road demos is contextual – it depends on the capability we want to highlight at that point. Many of our demos were done without using LiDAR altogether, using only the cameras. LiDAR is crucial for certain applications in autonomous driving, particularly those reliant on precise global positioning of vehicles. These applications necessitate accurate centimeter-level positioning relative to high-definition maps constructed using raw data. Our campus demos often use LiDAR as we showcase last-mile autonomy capabilities in such demos, requiring end effector constraints satisfaction (for example, as in autonomous parking). Our off-road autonomous driving demos also make use of LiDAR. However, we plan an off-road demo in August without LiDAR, relying solely on off-the-shelf cameras.

Our focus in autonomous driving technology development has shifted towards intelligence, specifically in decision-making and planning. Recent demos have showcased the effectiveness of perception-based systems, such as those using single forward-facing cameras, in navigating traffic and maintaining lane position. Previously, due to budget constraints, demos were conducted without LiDAR, emphasizing the importance of investing in decision-making capabilities rather than sensor technology. Ultimately, the choice of sensor is less significant than its ability to accurately represent the environment within acceptable tolerances for the given application.

Baidu and Waymo are already doing level 4 autonomous driving. How is your approach different from other companies?

Baidu and Waymo have adopted a traditional approach to autonomous driving, focusing on perception, mapping, localization, and planning algorithms. While visual sensors play a crucial role in creating a 3D representation of the vehicle’s surroundings, the emphasis lies in building high-definition maps and accurately localizing the vehicle within them. However, the decision-making and planning aspect of autonomous driving has often been overlooked by major players in the industry.

Despite advancements in perception and mapping, the level of sophistication required in decision-making remains challenging. The demonstration by Kyle Vogt (Cruise) showcasing a vehicle navigating through heavy traffic in San Francisco, even though impressive, primarily served as a validation rather than a display of true level 5 autonomy. The capability to handle unforeseen events and chaotic scenarios is where the true test lies. By integrating probabilistic reasoning into the planning algorithms, we aim to address the limitations of deterministic approaches commonly used in motion planning. Our demos, including the one conducted in February 2023 on campus and the bi-directional traffic negotiation on a single-lane road in October 2023, demonstrate our vehicle’s ability to analyze its surroundings and make decisions akin to human drivers without coming to a complete halt. While some companies like Tesla are exploring auto-regressive approaches, which require substantial resources, we believe our data-efficient reinforcement learning research and methodologies offer a promising alternative to achieving true autonomy.

What are you working on right now – in terms of technology and business?

Regarding our technology focus, pre-2021, explaining our approach was straightforward, but post-2021, it has become intricate. We have expanded beyond classical and non-classical frontiers into multiple branches. We are delving into areas like inverse reinforcement learning, tackling complex problems such as overtaking on 2-lane roads in India. Our bi-directional negotiation demo exemplifies our approach, evolving from classical reinforcement to deep reinforcement learning and enabling bi-directional negotiation at speeds of 80-100 km/h without LiDAR.

Securing funding has been essential, given the resource-intensive nature of our projects. Despite receiving $3 million in funding, we have spent half a million dollars, while other companies have received up to $100 million for similar endeavors. To further our goals, we aim to raise $12-15million in another seed round or directly go for $30-40 million in a pre-series round. This funding will facilitate the advancement and global expansion of our technology.

We plan to launch an Advanced Driver Assistance System (ADAS) tailored for Indian roads. Unlike existing systems, ours will operate beyond lane makers, enhancing safety for diverse road conditions. With investors interest, we aim to launch the ADAS system by year-end, marking a significant milestone in our journey.

We are also exploring defence applications and engaging with potential partners in that domain. Additionally, we’re considering developing our System on Chip (SOC) with retrofit capabilities for aftermarket use, expanding our market reach beyond OEM integration.

What is the current team size at the moment? What kind of skill set do you need?

Currently, we have 15 engineers. As predominantly an R&D company, our focus has been on developing new algorithms, prioritizing deep mathematical expertise coupled with proficiency in programming, particularly in C++ or Python. While programming and data skills were less crucial in the past, our progress in various domains now allows us to consider developers with extensive programming experience. However, historically, we have sought individuals with profound mathematical knowledge, often in theoretical computer science or mathematics, reflecting our emphasis on innovation in algorithm development.

Also read: In the fast lane – Autonomous driving technology