How Robotics Startup Sensori Moves from Onshape to NVIDIA Isaac Sim to Physical AI Training

How do you teach a humanoid robot how to perform a task on its own? To recognize an object, grasp it with the appropriate force, move it safely, and place it where it needs to go without human intervention?

Designing the robot is only part of the challenge. Developing the physical AI policy, which teaches that robot how to perceive its environment, make decisions, and safely interact with the physical world, introduces another level of complexity for robotics engineers.

Sensori, a Dallas-based startup, is taking on this challenge with Yuri, a wheel-based humanoid robot that can help researchers develop and test physical AI policies that would eventually power real-world robotic workforces.

Named for success (after the first human in space Yuri Gagarin), the robot has versatile grippers that can both handle larger objects and perform precise tasks, such as plugging in an Ethernet cable. Cameras in the head and wrists provide visual data for physical AI training, while force feedback gives both the teleoperator and the AI awareness of how much gripping force is being applied, helping the robot learn to manipulate objects effectively.

“The goal,” Sensori Co-founder and CEO Ross Melbourne says, is to have “full-blown robotic labor that’s doing work that is dull, dirty, and dangerous in warehouses and in factories.”

To get there, the Sensori team needs a suite of highly connected tools that can handle complexity and take Yuri from CAD model to a research lab to building physical AI policies. They chose PTC’s Onshape and the open robotics simulation framework NVIDIA Isaac Sim, using the SimReady Publisher, an Onshape Labs feature, which seamlessly imports the CAD model into the simulation environment.

The feature was developed by PTC using NVIDIA’s Omniverse libraries to create OpenUSD models with the necessary simulation-ready metadata, along with rich geometry, making them ready for simulation and reducing manual rebuilding between CAD design and simulation.

SimReady Publisher Collapses the CAD-to-Sim Gap

Before developing physical AI policies, a robot must be trained. Before a robot can be trained, engineers must understand how it will behave in the real world.

Traditionally, that process begins in CAD, moves into a simulation environment, and then continues through physical testing. This data handoff can require tedious rebuilding to ensure accurate joint physics and drive attributes. After testing, a design change might be necessary, and so the process repeats, dragging out each iteration.

With the SimReady Publisher, that process collapses. Much of the information can be defined in Onshape and exported directly as OpenUSD for use in NVIDIA Isaac Sim. As the design evolves, changes to the design can be pushed to Isaac Sim in a few clicks versus another manual rebuild.

Since OpenUSD contains the 3D model with the appropriate joints and configurations, “all you have to do is export it [from CAD], import it, and verify [in the sim tool],” explains Sensori robotics engineer Parker Sepulvado. “That’s a 5-minute process compared to what was originally a week process.”

“The way that Onshape designed mate connectors, rather than fixing everything in three-dimensional space, is very intuitive for robot joints,” he adds.

Streamlining Robotics from CAD Model to the Real World

With a validated robot model in NVIDIA Isaac Sim, Sensori can begin the process of teaching Yuri how to perform real-world tasks.

Using Telepath, the Sensori’s teleoperation and Physical AI software platform, an operator controls Yuri through a VR headset or body-worn control system. As the operator performs tasks such as picking up objects, moving materials, or interacting with equipment, the robot records those actions as training data. The system captures not only the robot’s movements, but also information such as visual inputs from cameras and force feedback from the robot’s grippers.

That data can then be used to create physical AI policies, the decision-making behaviors that allow a robot to perform a task autonomously. Rather than manually programming every movement, engineers can demonstrate a task, collect examples, and use those examples to train AI systems that can repeat the behavior on their own.

Simulation plays an important role in this process. Before hardware is built or modified, Sensori can use Isaac Sim to evaluate robot dimensions, stability, range of motion, and task viability in a physics-based environment. The closer the simulation matches reality, the faster the team can iterate and generate useful training scenarios.

“If you can teleoperate it in simulation in a physics-accurate way, then it’s highly likely that when you build it in reality, it’s going to work the same way,” Melbourne says.

For Sensori, the combination of Onshape, SimReady Publisher, Isaac Sim, and Telepath creates a connected workflow that moves from design and validation to data collection and AI training with fewer manual steps along the way.

Building the Foundation for Physical AI

Sensori joined the Onshape Startup Program while developing Yuri, giving the team access to cloud-native CAD tools well-suited for fast-moving robotics development.

As a startup, the company regularly iterates on robot designs, tests new concepts, and collaborates across disciplines spanning mechanical engineering, software development, AI, and controls. Maintaining a single source of truth for design data becomes increasingly important as those systems converge.

“As well as being an excellent technical tool that’s easy to use and is all well-connected, it has the sharing capability that [Parker] needed to work collaboratively,” Melbourne says.

The founders believe these kinds of connected engineering workflows will become increasingly important as physical AI moves from research environments into commercial deployments.

“America is on this new journey to bring manufacturing back to the United States,” Melbourne says. “[PTC] and NVIDIA and, hopefully, Sensori, will be a key part of that journey.”

For now, Yuri serves as both a research platform and a glimpse into that future – one where robots are designed, simulated, trained, and deployed through an increasingly connected digital workflow. As physical AI continues to advance, reducing the friction between design tools, simulation environments, and AI training systems may prove just as important as the robots themselves.