Getting Our Hands on Physical AI: What We're Learning in the Field

There's no shortage of videos showing humanoid robots doing impressive things. Walking across uneven terrain. Sorting packages. Folding laundry. The demos are polished; the movements fluid, and the implication is clear: physical AI has arrived.

We believe that's partly true, and partly misleading. The difference matters enormously to any organization seriously considering deploying robotics.

At WWT, advising clients on physical AI from a distance was never the goal. We wanted to get our hands on it. So, we did.

Solving a real problem on a real floor

We engaged with a large food manufacturer facing a challenge that's more common than it might sound. Their production lines shift frequently, sometimes daily, based on what's being made that shift. That kind of variability makes traditional fixed automation impractical.

At the center of each line sat a manual task that was tedious, physically demanding and hard to keep staffed: a worker scoops ingredients from a large container and transfers them into a hopper, which distributes the correct amounts into packaging moving along a conveyor belt. Standing in one spot, bending and scooping, all shift long. It's exactly the kind of repetitive, strain-inducing work that makes people think a robot should be able to handle it.

Our client thought so, too. When they came to us, they were excited about humanoid robots, specifically the bipedal, human-form-factor machines that have dominated the headlines. Part of our job was being honest with them about where the technology actually stands.

Separating hype from readiness

One of the first things we told this client was to temper expectations on robotic capabilities and timelines. Many of the demonstrations circulating right now take place in controlled environments, are teleoperated or are trained to showcase a specific capability rather than withstand the demands of a working plant.

Rather than promise an outcome we couldn't guarantee, we proposed something more valuable: a structured evaluation to understand what a given robot can and cannot do in this specific environment. What a robot can do is a different question from what it's ready to do at scale, and that distinction is where most organizations get into trouble.

The work we've been doing is focused on generating real evidence so our client can make an informed investment decision. It's a different kind of engagement than a deployment, but we think it's the right starting point for anyone serious about bringing robotics into their operations.

How to train a robot 

We conducted our evaluation with the Unitree G1, a humanoid robot about four feet tall with two arms, two legs, dexterous hands and various integrated sensors, including a RealSense depth camera. Running basic pre-programmed functions is straightforward. Getting it to perform a dynamic, repeatable task in a variable environment takes real work.

To get the Unitree G1 to perform the scooping task, we used teleoperation. A team member wears a Meta Quest VR headset that maps their arm movements one-to-one to the robot's arms in real time, so moving your arm moves the robot's arm, and flexing your fingers flexes the robot's fingers. From there, we attached a bucket to the robot's hand and built a demo environment with a container and packing peanuts to simulate the ingredient box, then recorded multiple episodes of the robot performing the scooping motion under human control.

Those recordings capture two things simultaneously: the robot's visual feed from its cameras and the precise joint positions at every moment during the task. That paired data (what the robot saw and what it did) becomes the training set for a machine learning model.

The model learns to associate visual inputs with appropriate physical actions. Rather than training a single monolithic model to handle the entire task, we're building separate models for distinct phases: one for the scooping motion, one for the dumping motion and another for lower-body navigation. Combining everything into a single model at this stage creates compounding complexity and pushes us into deep reinforcement learning territory before we've validated the simpler components. Breaking it into phases lets us isolate what's working and what isn't.

In parallel, another member of our team took a different approach entirely: training a policy for the same task using reinforcement learning inside a simulation environment, without any teleoperation data at all. Running both approaches side by side gives us a clearer read on where each technique holds up, rather than betting the whole evaluation on a single training method.

What the training data must get right

One of the more counterintuitive lessons from this work is how much the quality and variety of training data determine whether a model holds up outside the lab.

Training exclusively on recordings from a single angle, lighting condition and box texture produces a model that works well in that exact scenario but struggles when conditions shift. Floor surfaces that reflect light differently, subtle changes in ambient lighting, containers of a different color — these are standard features of any real manufacturing environment, and a model that hasn't encountered them in training will show it.

Knowing this, we deliberately vary the training data across scooping angles, positions and lighting conditions. We're also exploring synthetic data augmentation, using simulation tools to generate training scenarios with varied reflections, surfaces and light sources that would be impractical to recreate physically every time. The goal is a model robust enough to generalize to the natural variability of a real floor.

Safety planning starts early

Before any trained model runs on the actual robot, it runs in a simulation environment first, specifically to check that the robot won't behave erratically in ways that could injure a person or damage equipment. A model that starts flailing its arms unpredictably the moment it's deployed is one that never should have left simulation. 

The broader safety picture for a real deployment is even more complex. Any robot operating near people in a working plant needs safety systems that shut it down when a human enters its operating zone. The robot will eventually need to meet food-safety standards, whether through materials, cleaning protocols or protective measures. And the operational math of keeping a robot running across long shifts must work: on many platforms, current battery life doesn't cover a full production day, which means planning for opportunistic charging or rotating multiple units.

None of this is insurmountable, but it requires planning well before anyone talks about deployment.

Form factor and intelligence are not the same thing

Something that keeps coming up in our conversations with clients is worth stating plainly: humanoid form factor and AI-enabled intelligence are not the same thing.

Many organizations see videos of bipedal robots and assume the human shape is what makes them smart. The intelligence layer — the computer vision, the model training, the decision-making that lets a robot perceive its environment and act on what it sees — is entirely separable from the physical form factor. A highly capable, context-aware robotic arm can make smart, real-time decisions without looking anything like a person, and a humanoid platform is only as capable as the models running it.

This distinction matters for anyone evaluating robotics investments. The right form factor for a given task is the one that performs the task reliably, safely and cost-effectively.

For the scooping task we've been working on, the Unitree G1's bipedal design introduces complications worth noting. Because it's bipedal, the lower-body model has to account for balance as the robot lifts weight away from its center of mass, the same physics a human manages naturally, but one that adds meaningful computational load for a robot. Its weight tolerance is also limited, which matters when you're scooping ingredients that have real mass. Our evaluation of the humanoid form factor is honest and ongoing, and that's the point.

What we're taking forward

The work with the Unitree G1 has already yielded valuable insights into what evaluating and deploying physical AI requires. A few things we're carrying forward:

• Task decomposition beats end-to-end training, at least for now. Breaking the target operation into discrete, trainable subtasks is more tractable than training the whole behavior at once. The complexity compounds fast otherwise, which is part of why running teleoperation and simulation-based reinforcement learning in parallel was worth the extra effort.
• Training data quality matters more than volume. Varied, well-captured teleoperation data is the foundation on which the model is built. Shortcuts here show up later as brittleness.
• Simulation isn't optional. A model doesn't touch the real robot until it's proven safe in simulation first.
• Separate the hype from the hardware. The most useful thing we can do for clients right now is give them an honest picture of what's possible today versus what's on the roadmap, which means running the work ourselves rather than relying on vendor demos.

What this means for your organization

If you're in manufacturing, logistics, food production, warehousing or any environment where physical, repetitive tasks create staffing and safety challenges, robotic automation is worth exploring seriously. The technology is real, and it's advancing quickly. Organizations building hands-on experience with it now will have a meaningful head start.

The path from a demo that looks impressive to a robot running reliably on your floor requires task analysis, deliberate model training, safety planning and an honest evaluation of which form factor fits the job. That's the work we did with this client, and it's the work we'd do with yours.

WWT's AI and Data practice is actively engaged in physical AI research and deployment evaluation. To learn more about our robotics capabilities or to explore what this work could look like in your environment, reach out to your WWT account team.