Autonomous driving system maker, Imagry, Inc., promises to convert normal passenger cars into driverless vehicles by installing its artificial intelligence (AI)-powered systems for a fraction of the cost of a robo-taxi.
“When you’ve built a software and hardware stack that costs about $100,000 per vehicle, there is no business case for fitting it into a passenger vehicle that costs around $30,000. These expensive systems may make sense for robo-taxis that drive 12-14 hours per day because you eliminate the need for drivers, who account for a significant portion of the overall cost,” CEO of Imagry, Eran Ofir, told Mint in an interview from his Israel-based office. “In a passenger vehicle, though, you need something that justs costs $2,000 to $3,000 extra.”
Instead of depending on high-definition (HD) maps, also known as 3D maps that many car makers typically use in driverless cars, vehicles powered by Imagry’s AI systems interpret real-time visual information–such as objects, lanes, traffic lights, and pedestrians–to make driving decisions. Most driverless cars use HD maps work in conjuction with Light Detection and Ranging (LiDAR) systems and radars.
While the maps provide a static, detailed road environment, LiDARs and radars offer real-time, dynamic data on objects and obstacles. The sensor data provides real-time updates. Waymo’s autonomous vehicles, for example, use a combination of HD maps, LiDAR, and radar to navigate complex urban settings with high precision. However, Elon Musk-run Tesla and UK-based startup Wayve do not use external HD maps. Instead, they use camera-based vision, deep neural networks, and real-time sensor data to provide the cars with real-time learning like humans do.
HD map-based systems offer high precision, context, and improved predictability, but they’re expensive, require constant updates, and are vulnerable to cyberattacks. Mapless systems rely on real-time data from sensors, making them cost-effective, adaptable to real-time road changes, and scalable. But they may struggle with complex environments and lack predictive capabilities without prior map data.
The debate over the trade-offs in each approach notwithstanding, autonomous driving system maker, Imagry, Inc., is betting on the latter method. Its autonomous driving system for cars and even buses is fundamentally driven by AI, according to Ofir.
This “bio-inspired design” relies on two primary components: perception and motion planning. “Perception mimics human vision by interpreting objects and surroundings, while motion planning replicates cognitive processes occurring in the brain’s cortex,” Ofir explained.
In India, the object detection system must be trained with specific images, such as cows, to ensure it understands the necessity of yielding to them on the roads.
The system uses video feeds from multiple cameras, and processes this data through a series of “specialized deep convolutional neural networks (CNNs)”. A CNN is a deep learning (subset of machine learning, which is an AI technology) model that analyses visual data by recognising patterns and features in images. In driverless vehicles, CNNs process camera feeds to detect objects like pedestrians, traffic signs, and other vehicles.
Each network is trained on millions of images to detect and classify various objects, such as traffic lights, signs, lanes, pedestrians, and vehicles, covering a 360-degree view within a 300-meter radius, according to Ofir.
“This allows the system to construct a real-time, HD-equivalent (3D) map of the environment, which is then integrated into the motion planning component that learns and adapts over time through a process called learning by imitation—similar to how a child learns,” Ofir explained. He added this system can be integrated with LiDARs and radars, “if needed”.
Also Read: Meet the four musketeers of India’s quantum computing dream
According to Ofir, many robotic taxi services that use HD maps get confused “if road conditions deviate from their programmed expectations”. HD Maps also require constant high-bandwidth communication, which can pose challenges in areas with network congestion, rural locations, or tunnels, where signals may be weak or nonexistent, according to Ofir. He added that HD maps are costly to maintain and can expose vehicles to cybersecurity vulnerabilities.
View Full Image
Fortune Business Insights predicts that the global autonomous vehicle market size, which was valued at $1.5 trillion in 2022, is forecast to touch $13.63 trillion by 2030. Imagry, according to Ofir, has already signed a 10-year deal with Continental to develop a software-defined vehicle platform that incorporates their neural network-based motion planning, according to Ofir. This allows vehicles to gradually gain autonomous driving capabilities through over-the-air software updates. Starting with features like autonomous parking and traffic jam assistance, the system ultimately aims to achieve full self-driving functionality.
“Moreover, our system is hardware agnostic, allowing customers—including Tier 1 suppliers and OEMs (original equipment manufacturers)—to choose their preferred hardware without being tied to specific components,” Ofir said. He added that this hardware independence is particularly advantageous as it accommodates the diverse needs of various vehicle price segments–from entry-level to high-end models, which often have vastly different computing capabilities and sensor configurations.
Autonomous bus
From a safety perspective, Imagry is currently focusing on passenger cars with “Level 3 autonomy”, where the vehicle can drive itself, but the driver remains responsible. The company is also focusing on L4 driverless electric buses. These vehicles do not require human interaction in most circumstances, but a human operator can manually override the system for safety reasons, which could occur if the AI model makes mistakes or delays decisions.
Autonomous buses, according to Ofir, “are simpler to implement than passenger vehicles because they operate in geo-fenced areas and follow fixed routes, allowing for quick learning of the environment.” He added that despite the potential for faster deployment, stringent regulations govern the introduction of autonomous buses, including passing rigorous safety tests and cyber assessments before allowing passengers.
According to Ofir, there’s a global shortage of bus drivers—currently at 15% and increasing—which places pressure on public transportation operators. This shortage has led to pilots for autonomous buses in 22 markets worldwide, including Japan, where the government plans to deploy 50 autonomous bus locations by 2025 and 100 by 2027.
The economic implications for operators are significant, says Ofir, with estimates suggesting that autonomous buses could save between $40,000 and $70,000 annually per vehicle. “This could drastically improve operational margins and service levels. For instance, a partnership with a leading public transportation operator with a fleet of 70,000 buses exemplifies the transformative potential of this technology,” Ofir said.
Adaptability and customization
Imagry’s AI systems can also operate in diverse environments, including congested urban areas in Israel and Tokyo, “adapting to driving on either side of the road, a feat not commonly achieved by other autonomous driving companies”. For a vehicle to operate in a new area, the entire system requires time to learn the local traffic patterns and road characteristics, he explained.
While most traffic signs are similar across countries, unique signs or rules exist that need to be programmed into the system. For instance, when transitioning from the US to Germany, the software must be updated to reflect local rules, such as the prohibition of right turns on red.
While India does not have driverless cars, Ofir believes that the primary challenge of operating such cars “is not congestion, as the system can adapt to traffic jams, such as those found in Indian cities like Bengaluru or Pune”. Instead, the focus is on training the software to recognize and respond to local conditions. “For example, in India, the object detection system must be trained with specific images, such as cows, to ensure it understands the necessity of yielding to them on the roads. This targeted training helps the system adapt to unique driving environments effectively,” Ofir concluded.
