Autonomous systems operate through finely tuned perception models designed to process and respond to complex real-world inputs with accuracy. At the core of this intelligence is meticulously labelled data, and that's where iMerit, the company that brings together the right technology, skilled talent, and proven methods to deliver reliable AI data solutions, has quietly become indispensable. As a trusted partner in the high-stakes world of autonomous mobility, the company blends human expertise with cutting-edge technology to deliver pixel-perfect data annotation across vast multimodal inputs—from LiDAR and radar to high-resolution camera feeds. Every pedestrian, street sign, lane marking, and emergency vehicle the AI must learn to recognise has been precisely tagged by specialists at iMerit, Sudeep George, CTO of the company has said.
During an interaction with Mobility Outlook, he said, the company enables perception algorithms to train on millions of real-world scenarios, forming the very foundation on which self-driving systems operate. By combining automation with human-in-the-loop quality assurance, it helps perception AI not just see—but truly understand—its environment with the accuracy needed for real-world autonomy, he said.
Missed Inference
Vehicle perception is one of AI’s toughest frontiers. Unlike static environments, roads are endlessly dynamic—lighting shifts, obstacles appear suddenly, and human behaviours are unpredictable. A self-driving system must do more than spot objects; it must understand context, predict intent, and respond instantly. “A missed inference isn’t just a software glitch—it’s a safety risk,” said George. What comes naturally to human drivers becomes a complex challenge for AI.
Sudeep George
This challenge spans far beyond passenger cars. iMerit works across the autonomous spectrum—from delivery bots and tractors navigating rugged farmland to massive Class 8 trucks hauling freight on highways. In agriculture, even structured environments require machines to adapt to changing terrain and low visibility. On highways, heavy trucks—up to 80,000 pounds and over 70 feet long—demand perception systems that can handle high-speed decision-making with no room for error.
No matter the vehicle, the requirement is the same: terabytes of data, and each pixel matters. iMerit’s role is to bring clarity to that chaos—through precise data labelling, validation, and continuous feedback. “Helping AI see clearly is how we help make autonomy safer,” he said.
Perception in autonomous systems begins with cameras—but doesn’t end there. At iMerit, the focus is on full sensor fusion pipelines that combine cameras, LiDAR, and other inputs to build a richer, more reliable understanding of the environment. “Picture this: a camera picks up a partially obscured shape, while LiDAR adds depth and motion cues to confirm it’s a pedestrian crossing at an angle. By merging these signals, the AI makes a confident, safety-critical decision. This multimodal approach creates layered redundancy, where each sensor complements the others, reducing blind spots and strengthening real-world perception,” George explained.
Real-time Orchestration
Behind the seamless fusion of sensor data lies not traditional middleware, but a purpose-built edge computing system embedded directly in the vehicle. This isn’t just processing—it’s real-time orchestration. At highway speeds, milliseconds matter, and cloud offloading is too slow. Instead, the vehicle’s onboard AI stack takes raw inputs from LiDAR, radar, cameras, and ultrasonic sensors and transforms them into decisions within microseconds.
“This edge AI functions as the vehicle’s brain. It must spot a stop sign obscured by branches, anticipate if a cyclist will turn, or distinguish a shadow from actual debris. Unlike human drivers, it learns all this from scratch—through millions of annotated examples. Platforms like Ango Hub are critical here, curating and validating the data that trains and fine-tunes these intelligent driving systems,” he noted.
Understanding Human Sense
Humans rely on conceptual reasoning—we instantly recognise that a man in a Santa suit is still just a man. AI, on the other hand, learns through patterns. It doesn’t “understand” in the human sense; it needs millions of labelled examples to recognise variations. That’s where iMerit’s role becomes crucial. “The team continuously annotates, validates, and tests data,” especially edge cases that defy typical patterns. This constant exposure helps the model generalise better. Over time, it’s not just about feeding more data, but about identifying the rare exceptions that unlock the final layer of accuracy.
New Frontier
George said Generative AI is opening a new frontier in autonomous mobility—giving vehicles a voice. Instead of silent decisions or blinking lights, AI narration offers real-time updates like, “Approaching a stop sign… checking for cross traffic… turning left now.” It’s about trust, transparency, and making passengers feel safe and informed. Even more vital is how it handles uncertainty: “I’m unsure how to proceed. Please take over.” That moment of clarity can be life-saving.
This evolution, powered by large language models, bridges machine intelligence with human assurance. It demands a new kind of AI training—where perception, decision-making, and speech come together in complex, contextual ways. iMerit plays a pivotal role here in terms of tuning, validating, and stress-testing generative AI outputs, and designing simulations that push beyond pattern recognition into emergent decision-making. “It’s a shift from labelling pixels to guiding intelligence,” he said.
Challenges In Training
According to George, training autonomous systems isn’t just about feeding data—it’s about feeding the right data. The company tackles this with sharp focus on quality. One mislabelled object can trigger a cascade of errors, so “precision matters at every pixel.” But even perfect data can't account for every real-world surprise. That’s where exception handling comes in. The team builds intelligent guardrails to flag anomalies and feed expert insights back into the model.
The real challenge lies in managing complexity—edge cases, diverse environments, and large-scale annotation. Through continuous feedback and validation loops, the tech company ensures that perception systems not only learn, but evolve safely and reliably.
3D Sensor Fusion, 3D Point Cloud & LiDAR
Catastrophic Forgetting
One of the quiet threats in AI training is catastrophic forgetting—when a model, after learning something new, forgets what it previously knew. In autonomous driving, this can be dangerous. Picture a vehicle trained to navigate snow, then retrained for summer city traffic. Without careful tuning, it may perform well on dry roads but fail when snow returns.
This happens because neural networks update weights with new data, sometimes erasing older, vital patterns. iMerit prevents this by starting with diverse, curated datasets that span conditions and edge cases. Continual learning workflows help retain past knowledge while adding new inputs. Human-in-the-loop reviews act as safety nets, spotting regressions that machines might miss. Because AI evolves constantly, so does the work behind it.
In autonomous driving, there’s no room for “we’ll fix it in the next update.” Unlike consumer apps, this isn’t a space for trial runs. “There’s no such thing as partial autonomy at Level 5 —either the system works flawlessly, or it doesn’t belong on the road. Traditional AI has brought us far, enabling perception, prediction, and planning. But now, the bar is higher. The next frontier is AI that doesn’t just act, but explains,” he mentioned.
Picture a vehicle that not only recognises a stop sign but says, “Slowing down for a stop sign. Visibility is limited. Proceeding with caution.” This blend of generative and explainable AI is what iMerit is helping bring to life. Trust in autonomy isn’t built on automation alone—it’s built when systems can show their thinking. For iMerit, the mission goes beyond training models; it’s about shaping AI that performs with precision and communicates with accountability, George concluded.
Photo courtesy: iMerit
Also Read:
IBM’s Hybrid Blueprint Secures Future Of Gen AI In Automotive