This website uses cookies

Read our Privacy policy and Terms of use for more information.

On the concrete floor of a healthcare facility under construction, a Dusty Robotics layout robot was printing building plans, showing construction crews where walls, pipes and other features belonged. But there was a problem. With its printer mounted in the center of the machine, the robot couldn't reach the areas closest to walls and other obstructions.

The company was testing its FieldPrinter robot with Level 5 Inc., a metal framing and drywall contractor whose relationship with Dusty dates back to 2021. Its engineers had tried to anticipate the demands of a construction site, but the customer spotted something they had missed.

“What about all the stuff along the wall?” the customer asked. The question prompted a mechanical redesign. Dusty moved the printers to the robot's edges and developed a front section that could extend toward walls and pipes, letting it mark areas the original design couldn't reach.

Customer trials sent Dusty back to the drawing board so FieldPrinter could operate close to pipes and walls. Credit: Dusty Robotics

The trial was initially unpaid, but Level 5 was so impressed that it later sent Dusty a check. The company still displays it on its front desk.

The site has other plans

Dusty tried to bring the construction site into the office. The team swept up debris from building sites and scattered it across the office floor. They also wore hard hats and gloves while testing FieldPrinter, replicating the protective equipment used on construction sites.

But some conditions were impossible to reproduce. At an early deployment in San Francisco, the robot's Wi-Fi connection became unusably slow. The culprit was a neighboring 30-story residential building, where the many individual Wi-Fi networks created interference that the team hadn't encountered in its office tests.

“That was something that we would have never expected,” said Dusty Robotics co-founder Philipp Herget. “It doesn't happen in the office, and we had to design around it.”

Herget has also been surprised by how customers find new uses for the FieldPrinter. Beyond marking where construction elements belong, customers can print instructions directly onto the floor. For example, a customer might print a warning not to close a wall before an inspector has checked it.

Previously, such an instruction might have appeared only in a project schedule. Now it can be placed where the work is happening. Herget said Dusty takes ideas from individual customers and shares what it learns with others.

The food factor

At Chef Robotics, a food automation specialist, working with customers presents a different set of engineering challenges. Its robots must adapt not only to the requirements of individual production lines, but to the unpredictable physical properties of food itself.

On a production line at Cafe Spice, robotic arms fitted with scoops portion rice into meal trays alongside human workers. The process is designed for repetition, but individual portions vary in shape and consistency, and even slight temperature changes can affect how ingredients behave.

Chef Robotics robots scoop rice into food trays on the Cafe Spice production line. Credit: Chef Robotics

“Food is weird,” said Luis Rayas, chief technology officer at Chef Robotics. “It’s one of the weirdest things to manipulate.”

For robots, handling that variability reliably is the challenge. Chef’s answer isn't to build a new robot for every recipe, but to develop a versatile core system that can adapt to different tasks. Its software uses depth and color data to interpret what is in front of the robot, while deployment experience helps the company develop capabilities for new ingredients and applications.

When needed, Chef can swap in a different utensil at the end of the arm without changing the robot itself. The company can also adapt how its equipment integrates with a customer's production line, while keeping the underlying system largely standardized. Rayas describes the aim as developing technology that is “useful in a broad set of circumstances,” rather than a bespoke machine for every application.

Chef began deploying robots at Amy's Kitchen in 2022, initially at its Santa Rosa, California, facility. Chef's case study reported a 12% improvement in portion consistency, measured by the proportion of food deposits falling within the desired weight range. Rayas explained that achieving this precision involves a combination of depth and color sensing, careful calibration and training data.

But portion consistency is only part of the challenge. The robots must also operate without frequent human intervention.

Rayas said that if someone had to intervene “even 5% of the time,” it would “pretty much kill… all advantage” of using the robots. Individual deployments may require software adjustments, additional training data or a different utensil to achieve the necessary reliability.

The exchange of expertise also works in the other direction. Rayas said Chef sometimes identifies problems elsewhere on a customer's production line that affect the robots' performance. An upstream bottleneck, for example, can make the robots appear to be the source of a slowdown.

Experience on production lines is also informing Chef's longer-term development. The company uses NVIDIA’s robotics simulation platform Isaac Sim to test parts of its software stack. However, conventional simulation struggles to reproduce food, which can melt, stick, deform or change consistency.

To address that gap, Chef is developing its Food World Model, a neural network being trained to simulate how ingredients behave. It is intended to complement Isaac Sim, which remains useful for representing the wider production environment and testing software, rather than replace it.

Chef Robotics’ Food World Model simulates how ingredients behave. Credit: Chef Robotics

The company is using data collected by robots already working on production lines to develop the model. Rayas said this had helped Chef get the project underway, giving it access to real-world observations that would be difficult to reproduce in a laboratory.

During his time on customers' production lines, Rayas has found that customers have little interest in the technology for its own sake. Sophisticated robotics might get a company's attention, but the real test is what it can deliver.

Customers care about the impact and the results, he said. They want to know: “What can this thing actually do for me?” 

Did you like the article?