On October 20 I spent the day at Long Horizon, Encord’s physical AI summit at Fort Mason in San Francisco. Day one was about robotics and embodied AI: what it actually takes to get robots out of the lab and into production.
One highlight was a panel on humanoid robots, and the best part was that the speakers didn’t agree. Most panels end on “it’s early, but exciting.” This one gave three different answers to the question every operator is asking: how do you get a robot to do real work, reliably, on a real floor?
One speaker had just spent a week with a humanoid robot in a car factory. The robot runs on three AI brains working together:
One brain works out what needs to be done and in what order. It can take a moment to think, because the plan only changes when the job changes.
One decides how to move, like whether to use one hand or two. This has to happen in real time, part by part.
One keeps the robot balanced and safe. This layer can never wait on the others, because a robot that loses its balance next to people is not a productivity problem, it’s a safety one.
Splitting the work this way is the interesting design choice. Planning can be slow and thoughtful. Movement has to be fast. Balance has to be instant. One model trying to do all three would be too slow for some tasks and too risky for others.
Anyone who has built enterprise software will recognize the pattern. You don’t run your rating engine, your tendering and your audit trail in one process with one set of rules. You give each job the speed and the safeguards it needs, and you make them talk to each other cleanly.
The big change is how you give the robot new work. Before, engineers had to write detailed instructions for every situation. Now they explain the job in plain English.
And when the robot makes a mistake, anyone on the factory floor can help retrain it, not just a few experts.
That point is underrated. In most factories and warehouses, the people who know the job best are on the floor, not in engineering. If they can correct the robot directly, the system improves at the speed of operations instead of the speed of a ticket queue.
Another speaker pushed back. Factory engineers get nervous about a robot that works most of the time but fails in ways no one can predict.
His approach is different. Let AI write the robot’s code and test it in a virtual factory first, then hand engineers code they can read and check before it runs. He said big manufacturers like it for exactly that reason.
The argument is about predictability. An engineer who signs off on equipment needs to know what it will do, not what it usually does. A model that learns on the job is powerful, but it is hard to certify. Code that a person can read, test and approve is slower to change, but it is easy to trust.
The third speaker said both ideas can work together. Smart AI can plan the work, faster AI can handle the movement, and robots keep getting better on the job.
In other words, you don’t have to choose between flexibility and control. You decide which layer gets which. Let AI handle the parts where flexibility pays off, and keep the parts where a mistake is expensive readable, testable and approved by a person.
When will a humanoid robot work a full 8-hour shift? Answers ranged from next year to a joking 25 to 100 years.
I come at this from transportation, and robots are already on the dock. DHL has signed for more than 1,000 additional Boston Dynamics Stretch robots, each able to unload trailers at up to 700 cases an hour. In the 2025 MHI and Peerless Research Group study, 48% of respondents said they already use robots, up from 23% in 2022, and 53% of respondents worked for companies under $50 million in revenue. This is not only a big-company story.
The panel’s three views map almost one-to-one onto how AI is entering the transportation plan:
The same shift is happening at the planning desk. In Zoree TMS, a planner can ask our built-in assistant to rate a lane, tender a load or update an order in plain English, and it carries the action out.
Every action the assistant takes shows a confirmation card that a person approves before anything changes. Flexible where it helps, controlled where a mistake costs money.
When a person, a carrier system, a warehouse robot and an AI agent can all change the same shipment, you need to know which one did what. Zoree records every change at field level, with who or what made it and when.
A robot that unloads 700 cases an hour is still idle if the trailer shows up three hours late, at the wrong door, or with nobody told it was coming. Robots speed up execution. They don’t decide what moves, when, on which carrier, or into which door. The transportation plan decides that.
That is why I left the panel thinking less about the robots and more about the systems around them. Whichever view wins on the factory floor, the plan feeding the robot has to be just as fast and just as trustworthy.
Zoree builds AI-native transportation software with a person in the loop for every decision that matters. We’d like to hear how your operation is approaching it.