Robots are getting more capable, but the next challenge is giving them enough experience in the environments where they will actually work.
GXO runs warehouse trials of humanoid robots from Reflex Robotics.
Robotics is moving incredibly quickly. New foundation models, humanoid robots, manipulation systems, and autonomous machines are demonstrating capabilities that would have seemed unrealistic only a few years ago. But getting these systems to work reliably outside of controlled environments is a different challenge.
The next step is getting more robots into the real world. Warehouses, factories, hotels, construction sites, machine shops, restaurants, and other businesses offer the environments and everyday physical tasks that robots ultimately need to understand. Finding ways to connect these businesses with the companies developing physical AI could help accelerate that transition.
Training physical AI requires more than teaching a robot how to complete a task once. Robots need experience with different objects, layouts, people, equipment, lighting conditions, interruptions, and all of the small variations that happen naturally when work is being performed.
Collecting that experience at scale is difficult and expensive. Labs and simulation can provide enormous amounts of useful training, but real workplaces contain complexity that is difficult to reproduce. Creating more opportunities for robots to operate in these environments could provide the diverse data needed to make physical AI more capable and reliable.
Almost every business that involves physical work contains tasks that could become valuable environments for robot training. Warehouses move and sort products, hotels restock rooms and handle supplies, machine shops load and organize parts, restaurants move ingredients and dishes, and construction sites constantly move tools and materials.
These environments offer something difficult to reproduce in a lab: real workflows with real variation. The tasks do not need to be especially complex to be useful. Repetitive activities involving movement, object handling, organization, inspection, or interaction with existing equipment can provide opportunities for robots to gain experience while developers learn what their systems can and cannot yet handle.
The most interesting model may be one where collecting robot data and creating value for a business happen at the same time. Instead of deploying robots solely to generate training data, they can be introduced around tasks that businesses already need performed.
Early systems may still require supervision, intervention, or assistance, but those interactions can become part of the learning process. As the technology improves, the same deployments can potentially transition from primarily providing valuable training experience to performing increasingly useful work for the businesses hosting them.
Robotics companies need access to environments where their systems can encounter real tasks, while businesses may be interested in automation without knowing which technologies are ready to explore. Creating stronger connections between these two sides could make it easier to move promising robotic systems into real operating environments.
This could begin with something as simple as visiting a business, understanding how work is currently performed, and identifying physical tasks that might be relevant to a particular robotic system. From there, developers and operators can explore small deployments that provide useful feedback before expanding into more ambitious applications.
We recently started the Real-World Robot Training Data project to explore this model across a broad range of industries. The goal is to build a network of businesses interested in hosting robotic systems for real-world training, testing, and deployment, and connect them with robotics and physical AI companies looking for those opportunities.
A warehouse might provide completely different training opportunities than a hotel, machine shop, construction site, or restaurant. Bringing together a diverse group of environments could make it possible to match different robotic systems with the places and tasks that are most useful for their development.
We’re interested in connecting with robotics and physical AI companies looking for real-world environments, as well as businesses curious about how emerging robotic systems could fit into their operations.
If you’re developing robotics technology, operating a business with physical workflows, or simply have an environment that could provide interesting opportunities for robot training and deployment, we’d be interested in hearing what you’re working on and exploring what might be possible.
OpenDroids is developing and deploying adaptable robotic systems for hospitality, logistics, manufacturing, retail, and other industries, combining custom hardware, AI, integration, and Robotics-as-a-Service.
Service Robot Co. helps businesses find, integrate, and deploy commercial robots for cleaning, delivery, material handling, manufacturing, and other real-world applications.
Chef Robotics uses AI-powered robotic systems to automate food portioning and assembly, helping food manufacturers increase production, reduce labor costs, and improve consistency across high-volume operations.
New reinforcement learning research from Physical Intelligence shows how robots can improve precise manipulation skills using minutes or hours of real-world experience.
Creating real-world deployment opportunities where robots can perform practical tasks and generate valuable training data for physical AI.
Developing highly automated food production facilities that combine robotics, industry expertise, and regional partnerships to reduce costs and supply prepared food at commercial scale.
CivCheck uses AI to review building plans for code requirements, helping applicants catch issues earlier and cities process permits faster.
Trunk Tools brings artificial intelligence directly into construction workflows, helping project teams find information, review documents, track changes, and work more efficiently with the project data they already use.
