A report from Deloitte and The Manufacturing Institute cites that the manufacturing industry will need 3.8 million labor roles filled by 2033. As of 2024, it is projected that roughly half of those roles will remain vacant.
Some of the positions with the widest gap include machine tending and bin picking. These are precisely the tasks today’s robotics platforms are fit to automate with Physical AI.
In this excerpt from our Top 10 FAQs guide, we detail some of the fundamentals of Physical AI deployment, including what it is, the deployment paths, and why this automation technology is important for manufacturers today.
1. What is Physical AI?
Physical AI is artificial intelligence that’s embodied in machines to perceive, reason about, and act on the real, physical world. The Physical AI vs traditional automation conversation is largely about its ability to handle unstructured tasks.
Traditionally, unstructured tasks such as picking varied objects from a bin were considered too complex, requiring costly reprogramming and overhauls. Conventional robots need parts fixed in precise positions to ensure that they know what to expect at all times; Physical AI systems can handle uncertainty, detecting randomly placed parts, figuring out where to grip them, and planning a collision-free path on their own. They can also handle retry attempts.
In short, Physical AI is the convergence of AI perception/decision-making with robotics hardware, motion planning, and edge compute, enabling machines to autonomously handle real-world, unstructured physical tasks.
2. What Are the Most Common Pathways to Deploy Physical AI?
In the last couple of years, attempts to solve the performance and dexterity problem have coalesced into two distinct camps for applied Physical AI solutions.
On one side, we have AI Pipelines, which treat robotics as a geometry problem. On the other side, we have End-to-End Models, which treat robotics as a learning problem:
- AI Pipelines: This architecture combines computer vision with classical/rule-based control. Better suited for tasks with visual uncertainty but structured execution, such as robotic bin picking or guided assembly, where deterministic motion follows once objects are detected; they offer more predictable cycle times and easier safety certification.
AI Pipelines today fit the use case for manufacturing better since they need less data, which makes them more efficient and economical. As AI Pipelines continue to evolve, adopting sensor-to-action models, they will offer greater dexterity and ability to handle complex interactions.
- End-to-end / VLA (Vision-Language-Action) models: These models excel at tasks requiring real-time adaptive decision-making and high dexterity, such as bi-manual manipulation or handling deformable objects, where both perception and action must adjust dynamically to unpredictable conditions.
3. Why is Physical AI in Manufacturing Important Now?
Physical AI is more relevant than ever before as three factors converge.
Labor Shortage: According to Deloitte and the Manufacturing Institute, the sector will need 3.8 million new workers between 2024 and 2033, but roughly 1.9 million of those jobs could go unfilled, a 50% shortfall, with retirements making up the largest share. As those experienced workers leave, they take decades of hands-on skill with them, and they leave behind exactly the dexterous, high-variance jobs like bin picking, machine tending, and palletizing that traditional automation could never handle.
Technology Maturity: Advances in foundation models mean Physical AI systems now handle these unstructured tasks closer to out of the box, instead of requiring weeks of data gathering and retraining every time the environment changes.
ROI: Physical AI ROI now lands within about two years, and per the World Economic Forum, in the near future Physical AI can cut engineering effort by up to 70% and accelerate time-to-value by up to 50%. The shortage was always coming. What’s new is that there is finally a way to automate the work that’s hardest to staff.
Get the extended e-guide edition of this article that dives deeper into Physical AI and includes 7 additional frequently asked questions.
