#45 Kary Bheemaiah: Beyond The Robotics Hype - The Hard Truth About Automation & Role Of Edge AI

In this talk

1. 4:19

   Edge AI brings computing closer to the work

   Bheemaiah describes edge AI as running smaller, task-specific models near the equipment that generates the data. This can reduce delays, cloud data-transfer costs, and concerns about sharing sensitive factory information, while still allowing useful results to go to the cloud.

2. 11:41

   Integration can outweigh the robot’s price

   In existing factories, robots must coordinate with systems such as PLCs, SCADA, MES, and ERP. Bheemaiah says integration can cost six to twelve times the robot’s purchase price, driven by protocol mismatches, manual data mapping, synchronization, and testing. He sees edge models as a possible way to translate between systems, though factory deployments still need careful trials.

3. 20:25

   Focus on one use case and learn from failures

   Bheemaiah points to a shortage of robotics training data and the variety of physical tasks as barriers to general-purpose systems. He recommends choosing a specific, repeated problem, such as pallet wrapping, and making the first customers similar enough to reveal recurring failures. He says those lessons can cut deployment costs over time; he also urges teams to plan for downtime from the start.

4. 30:09

   Assess readiness before automating

   Bheemaiah’s ARCHITECT framework starts with an honest assessment of a facility’s current capabilities, particularly because many manufacturers operate older systems. Teams should build missing foundations, select use cases based on return or operating cost, and account for uptime: he cites an automotive downtime estimate of about $250,000 per hour. He also stresses involving workers and making progress through successive improvements.

5. 42:47

   Training and human oversight remain essential

   Bheemaiah says teleoperation can help teach movement and force, and let a person intervene when a deployed robot gets stuck or confused. Depending on the job, training may also use motion capture, digital twins, simulated environments, and first-person video. He argues that each pipeline should match the task and its available data.