The robotics industry has fixated on a false problem. While researchers and startups chase larger datasets and more sophisticated AI models, manufacturers struggle with something far more mundane: the months of engineering work required to integrate vision systems into factory floors.
Rudy Cohen, CEO of Inbolt Physical AI, plans to challenge this narrative at RoboBusiness next month. According to The Robot Report, Cohen will present a talk titled "Physical AI Doesn't Have a Data Problem; It Has a Deployment Problem" on October 20 in Santa Clara, California. His argument rests on operational evidence from over 100 factory deployments spanning 40 million individual robot cycles.
The Real Constraint in Manufacturing
Inbolt's technology enables robotic arms to grasp components with flexible positioning, then uses real-time in-hand localization to correct trajectories mid-motion. This approach has proven effective across automotive suppliers and major manufacturers including Stellantis, Toyota, and Ford. Yet despite this commercial success, Cohen contends that the industry has misidentified the fundamental limitation holding back broader adoption.
The prevailing wisdom suggests that physical AI systems need better training data, more carefully curated datasets, and improved simulation environments. These factors may constrain academic progress and isolated laboratory demonstrations. However, Cohen's field experience points elsewhere: the friction lies in the loop connecting perception systems to motion control and the sprawling integration overhead that accompanies every factory installation.
"Physical AI is already commercially proven, just not where the headlines are looking." This perspective challenges the narrative that dominates venture capital and media coverage in the robotics sector.
Deployment as the Hidden Tax
Factory integration remains a six-figure undertaking for each deployment, Cohen suggests. This cost stems not from algorithmic limitations but from the fixtures, wiring, software configuration, and testing cycles required to adapt generic robotic systems to specific production environments. Closing the perception-to-action loop at servo frequency remains technically possible but operationally burdensome.
Inbolt's approach involves building a real-time control layer that transforms digital twins into executable robot commands, effectively compressing deployment timelines and reducing integration costs. Since its 2019 founding, the company has raised approximately 23.2 million dollars and expanded operations across Europe, the United States, and Japan.
Market Implications
Cohen's RoboBusiness presentation signals a potential shift in how the physical AI community evaluates progress. If deployment complexity rather than algorithmic capability represents the true constraint, then competitive advantage accrues to companies that streamline integration rather than those that accumulate training data or scale model parameters.
This framing carries implications for venture funding priorities, research direction, and where manufacturers should focus their robotics investments. The distinction between theoretical capability and practical deployment has long plagued emerging technologies, but Cohen's extensive factory experience provides concrete evidence for the robotics sector.
