Why commercial cleaning could be the first real-world test for physical AI | Talk w/ Gatlin Robotics’ CTO

Robotics is often associated with warehouse automation, self-driving cars, or humanoids doing housework. Gatlin Robotics believes commercial cleaning could be one of the earliest practical applications of physical AI.

The reason is simple: cleaning requires more than navigation. Tasks like picking up trash, moving chairs, opening doors, and wiping surfaces depend on manipulation, which is one of robotics' biggest challenges. While most autonomous cleaning robots focus on floors, Gatlin is building robots for everything else.

Offices, hotels, and similar spaces provide a useful balance of structure and variability. They are predictable enough for reliable operation, yet constantly changing as objects move and environments shift. As CTO Zach Vinegar explained on the Knoxlabs podcast, this makes commercial cleaning an ideal training ground for physical AI.

Because cleaning tasks are recoverable, every deployment also becomes an opportunity to improve. If a robot misses a task, a human can intervene, and that interaction becomes training data. Gatlin combines autonomy, teleoperation, simulation, and human supervision to gradually reduce the need for intervention.

Why cleaning is a useful test bed for physical AI

• The work is manipulation-heavy. Trash pickup, room reset, chair movement, doors, and surface wiping all require contact with the real world.
• The environments are structured but not static. Offices and hotels have repeatable layouts, but objects still move every day.
• Errors are recoverable. If a robot misses a task, a human operator can step in, complete the work, and turn the correction into training data.
• The business case is clear. Facility teams already understand cleaning schedules, labor pressure, service levels, and recurring monthly costs.

For Knoxlabs, Gatlin is also a good example of why robotics teams increasingly need an XR and teleoperation hardware layer. Headsets, motion trackers, haptic gloves, and body-tracking systems are becoming part of the robotics stack, not just VR demos. See the related Gatlin Robotics product page on Knoxlabs.

Training robots before they reach the customer

Before a robot is deployed, Gatlin Robotics builds a digital twin of the customer's facility.

According to Zach, the process starts with a short walkthrough of the space. That recording is converted into a 3D digital environment, where the robot can practice its tasks before arriving on-site. Instead of learning in a customer's office or hotel, the robot first learns in simulation. By the time the robot reaches the facility, much of the initial setup has already been completed.

Gatlin uses NVIDIA Isaac Sim to run these simulations. The company has also shared examples of converting office and conference room walkthroughs into USD environments, where robots practice navigation, light trash pickup, and chair manipulation before deployment. The more preparation that happens in simulation, the faster a customer can begin using the system.

This workflow also fits the company's Robot-as-a-Service model. Instead of delivering hardware and leaving the customer to configure it, Gatlin prepares the deployment in advance, then provides the robot, software platform, maintenance, fleet management, and teleoperation as part of a monthly subscription.

Why Gatlin is not betting on a single robot

"I think right now I'm very bullish on mobile manipulators." That's how Zach summarized Gatlin Robotics' current approach.

Humanoid robots receive most of the attention, but Gatlin is taking a more practical path. The company is already working with multiple robot platforms, including its own mobile manipulators, the Unitree G1 humanoid, and a quadruped equipped with a robotic arm. Each platform is chosen for the job it needs to perform, rather than trying to make a single robot fit every use case.

For commercial cleaning, Zach argues that mobile manipulators often make more sense. Offices and hotels are already wheelchair accessible, which means a wheeled robot can reach the places it needs to go while offering a simpler and safer platform. Humanoids become more valuable in environments with stairs, uneven terrain, or tasks that require greater mobility.

This philosophy extends to the software. Instead of building around a single robot, Gatlin is developing a robot-agnostic platform. As Zach explained, hardware is improving rapidly, with new systems appearing every few months. By keeping the software independent of any single form factor, the company can adopt new hardware as it becomes available rather than redesigning its entire stack.

That means the same fleet management and teleoperation platform can coordinate humanoids, mobile manipulators, quadrupeds, and future robotic systems through a single software layer.

Watch: The Future of Facility Cleaning

This Gatlin Robotics channel video gives additional context for how the company frames facility cleaning as a robotics service, not a one-off robot purchase.

Why touch matters as much as vision

One of the more technical parts of the conversation focused on a challenge that is easy to overlook: robots do not always have a clear view of what they are doing.

As Zach explained, a robot can see an object before reaching for it, but once its arm moves in front of the camera, vision alone is no longer enough. The robot still needs to know whether it has made contact, how much force to apply, and whether it is gripping or pushing the object correctly.

That becomes especially important for tasks like wiping a surface or picking up trash. Applying too little pressure may leave the task unfinished, while too much can damage the object or the environment. Vision can show where an object is, but it cannot always tell the robot how it feels to interact with it.

This is where force feedback comes in. As Zach explained, "you really use your sense of touch" when vision is limited. For example, when a robot's arm blocks its own camera view, touch can provide information about whether it has made contact and how much force it is applying.

By adding touch as another source of information, robots can better understand how they are interacting with the world rather than relying on cameras alone.

Better robots start with better data

One of the biggest shifts Zach described is how robots are trained today compared to just a few years ago.

Looking back at his time building burger-flipping robots at Miso Robotics, he explained:

"Everything was like a bunch of if statements... sort of like hardcoded positions of things. Nowadays it's very much like collecting the data, have it learn from that, and then generalize."

That idea runs through Gatlin's entire workflow. Digital twins, teleoperation, simulation, egocentric video, force feedback, and XR are all different ways of collecting data that robots can learn from.

As Zach put it:

"The real thing is do you have quality data... the data set is the real code now."

It is also why the company is hardware-agnostic when it comes to teleoperation. The team uses Pico and Meta Quest headsets and is evaluating MANUS gloves and force feedback to improve the quality of demonstrations. The hardware matters, but only if it helps capture better data.

Final thoughts

Cleaning may not be the first industry people associate with advanced robotics, but it brings together many of the challenges that define physical AI today: reliable manipulation, real-world data collection, simulation, and continuous learning.

Gatlin Robotics' approach reflects a broader shift in robotics, which is moving away from building machines that simply follow programmed instructions and toward systems that improve through experience. By starting with a specific, practical application, the company is using real deployments as a way to build more capable robots over time.

Related: Explore the Gatlin Robotics quote-only deployment page on Knoxlabs for the robotics, teleoperation, and XR hardware context behind the conversation.