Understanding Physical AI
How mobile robots learn inside virtual facilities
A study shows what training in a digital facility improves and why the decisive test still takes place on the real robot.
A mobile robot needs to cross a facility and reach its destination without getting stuck at obstacles. Its navigation strategy can first be trained in a virtual environment. Whether that strategy works on the real floor is a separate question. Researchers from Idealworks and NVIDIA investigated precisely this transfer in a study published in October 2026.
The team used a digital replica of Idealworks’ development centre, then tested the learned navigation on a real iw.hub. Adapting to this environment, including additional image variants generated with NVIDIA Cosmos, increased the success rate across four scenarios from 65.3 to 91.7 percent. The existing production navigation system achieved 100 percent. The study dated 24 October 2026 is a preprint; its figures apply to this experimental setup.
The result illustrates two points: virtual training can prepare a real task, and the transfer to the actual environment must be measured. A good simulator result alone does not establish how reliably a vehicle will reach its destination inside a facility.
A training environment with adjustable conditions
Consider a simplified example created for this article. A mobile robot must transport a container from a staging area to a machine. Between them are a straight aisle, a bend and a parked trolley. These elements can be arranged in a virtual model without rearranging the real facility for every trial. The robot’s dimensions also matter: a route clear for a small trolley may be too narrow for a wider robot.
The simulation calculates how the modelled objects behave and generates observations for the system being trained. A virtual camera image might show the aisle from the vehicle’s perspective. An action changes its position, which in turn changes the next observation. This creates a sequence of situations, decisions and outcomes.
A learning method can specify which behaviour is desirable. In our example, that could mean reaching the destination while avoiding collisions. Rewarding speed alone would be inadequate if the robot were allowed to brush against obstacles. The learning objective therefore needs to express what makes an acceptable operation. This relationship is fundamental to learning through trial and feedback, which NVIDIA also describes in its robotics research.
A single, unchanging virtual facility would, however, provide a very narrow training world. Lighting would always be identical, the trolley would remain in the same place and surfaces would look the same. A model could perform extremely well under those conditions without being prepared for changes encountered during operation.
Training therefore often varies properties of the environment. The technical term is domain randomization. NVIDIA explains the principle in its practical documentation: rather than seeking one perfect match with reality, parameters are varied within a range. The aim is for the model to cope with different conditions.
For the illustrative trip to the machine, the trolley might stand on the left of the aisle in one trial and on the right in another. Lighting could be brighter or darker. Such variations need to fit the problem being studied. A thousand images with different wall colours offer limited help if the real difficulty is a tight bend. The number of simulated trials cannot replace careful selection of situations.
The decisive comparison happens in the real facility
A virtual camera and a real camera do not automatically produce identical images. Movement, delays and contact with the floor can also differ from the model. This difference between simulation and reality is known as the sim-to-real gap. A digital replica is therefore a working model based on assumptions whose accuracy must be sufficient for the task.
A useful comparison for our example would test the same route virtually and then physically. Start and destination, obstacles and the success criterion would be defined. If a version worked only in simulation, the team could examine where its real behaviour diverged. That is more informative than comparing two entirely different routes under the same heading.
Several observations belong in the evaluation. Reaching the destination comes first, followed by travel time, stops and human assistance. A robot that arrives after an intervention should remain identifiable in the statistics. Likewise, a vehicle that regularly waits before a narrow passage may have a high completion rate while still delaying material flow.
A fictional test series illustrates the denominator: 36 successful arrivals out of 40 defined journeys represent a success rate of 90 percent. That describes those 40 journeys. It proves neither that every other route works equally well nor that the same rate will hold across an entire shift. Those claims require additional, appropriately designed trials.
In the study, the learned variants also experienced collisions and protective stops in more difficult scenarios. The production navigation system remained collision-free. The comparison also used different conditions: learned navigation processed camera images, while the production system used an existing map and laser sensors. The improved success rate alone therefore establishes neither equivalent safety nor a replacement for the existing system.
For businesses, simulation is useful for targeted preliminary trials. Bottlenecks, scenarios and training variations can be prepared before physical testing expands. A deployment decision subsequently needs results from the intended environment. Virtual training is valuable when its contribution can be demonstrated through a concrete, repeatable journey on the real system.
Sources checked: 12 November 2026. Research figures refer to the linked preprint. The route, proposed test design and calculation involving 40 journeys are illustrative examples created for this article.