In this episode of An Hour of Innovation, host Vit Lyoshin and Achyut Boggaram, a Senior Machine Learning Engineer at Torc Robotics, explore what truly goes on behind the scenes of autonomous trucks and why full self-driving has taken far longer than public timelines promised.They explore why autonomous trucks are not just an AI problem, but a safety-critical engineering challenge involving hardware, software, data, and regulation. The conversation explores how machine learning models interpret the real world, why edge cases are hazardous, and how autonomous vehicles generate massive amounts of sensor data in a matter of minutes. Achyut explains why redundancy, certification, and testing are treated more like rocket engineering than traditional software development. They also unpack common misconceptions about AI capability, data scale, and why impressive demos rarely reflect real-world autonomy.Achyut Boggaram is a senior machine learning engineer focused on applied AI research for autonomous trucking. He has led work on large-scale perception models, sensor fusion systems, and production machine learning pipelines that run directly on self-driving trucks. His expertise spans safety-critical AI, data infrastructure, and real-world deployment, making his insights essential to understanding why autonomy remains so challenging.Takeaways* A single missed annotation, like a stop sign or yield sign, can lead to catastrophic outcomes with an 80,000-pound vehicle.* Self-driving demos work in controlled environments, but real autonomy breaks down once conditions are unpredictable and unstructured.* Autonomous trucks can generate 600–800 terabytes of data in just 20 minutes due to raw, uncompressed sensor capture.* Machine learning models struggle to generalize the way humans do, even after billions of miles of training data.* Safety in autonomous trucking is treated like rocket engineering, with redundancy required at every hardware and software layer.* Autonomous trucks must run entirely on board without internet access, making real-time decision-making far more constrained.* When AI is uncertain, the safest response is not intelligence but a minimum risk maneuver, often pulling over or stopping.* Synthetic and photorealistic simulated data are now essential to train for rare but dangerous scenarios that may never occur in real life.* Autonomous systems can outperform humans in extreme conditions, detecting pedestrians at long distances in fog or darkness.* Autonomous trucks are not replacing drivers today, but filling a growing labor gap that could reach hundreds of thousands of unfilled jobs.Timestamps00:00 Introduction02:41 Why Autonomous Vehicles Still Struggle in the Real World05:40 What It Really Takes to Put Autonomous Trucks on Public Roads10:05 Safety Certifications That Decide If Autonomous Trucks Are Allowed15:50 How Self-Driving Trucks Generate Massive Amounts of Data20:09 How Autonomous Trucks Handle Dangerous and Unexpected Situations23:20 The Full AI Training Pipeline for Autonomous Vehicles31:33 The Most Critical Safety Gates in Autonomous Truck Testing34:21 Breakthrough AI Techniques for Fog, Night, and Extreme Conditions38:07 The Real Timeline for Autonomous Trucks Becoming Reality39:52 The Hardest Problems Blocking Full Self-Driving41:28 Are Autonomous Vehicles Inevitable?42:34 Electric vs Diesel Autonomous Trucks43:53 Will Autonomous Trucks Replace Human Drivers?48:09 Innovation Q&ASupport This Podcast* To support our work, please check out our sponsors and get discounts: https://www.anhourofinnovation.com/sponsors/Connect with Achyut* Website: https://torc.ai/ * LinkedIn: https://www.linkedin.com/in/achyutsarma/ Connect with Vit* Substuck: https://substack.com/@vitlyoshin * LinkedIn: https://www.linkedin.com/in/vit-lyoshin/ * Podcast: https://www.anhourofinnovation.com/

Ask this episode

“What did the council decide on item 42?” — answered with the speaker and the moment.

Transcript

Related