> ## Content Index
> Fetch the complete content index at: https://www.healthdatacon.io/llms.txt
> Use this file to discover other available public pages before exploring further.

# Johns Hopkins tests surgical robots in simulated operating rooms before moving them into physical ones
- URL: https://www.healthdatacon.io/articles/johns-hopkins-surgical-robot-simulation-hlth-2026/
- Published: 2026-11-18T17:35:00.000Z
- Updated: 2026-11-19T11:00:35.000Z
- Description: Computer scientist Mathias Unberath showed how digital replicas can test a robot’s ability to handle unfamiliar situations, and described where he would start with supervised automation.
- Author: HDM Staff
- Tags: AI, Americas, HLTH

Johns Hopkins researcher Mathias Unberath showed a robot attempting a room-turnover task in a mock operating room after its software had been tried in a digital replica. He used the example to show how researchers can expose a robot to many variations of an operating room before choosing a version of the software to test on physical hardware.

Unberath, a Johns Hopkins [computer science professor](https://engineering.jhu.edu/faculty/mathias-unberath/?ref=healthdatacon.io), presented the work at The Futurist’s [Physical AI Summit](https://thefuturist.global/PhysicalAI?ref=healthdatacon.io) during HLTH USA 2026 in Las Vegas. Summit host Ian Khan introduced the session, “From Simulation to the Bedside.” Unberath described what testing must establish as systems take on unfamiliar surroundings and choose their own next action. The room-turnover demonstration took place in a research setting.

What changes with greater autonomy is the evidence that a hospital or regulator would need. A machine executing a fixed plan can be assessed in part by how precisely and consistently it follows that plan. A system choosing a move in response to its surroundings also needs to be tested on the range of conditions in which it will make those choices. That carries implications for the researchers who design the tests and for the hospitals that may eventually deploy the equipment.

## A simulated room makes variation easier to test

Unberath described producing a three-dimensional counterpart to an operating room from recorded observations of the real environment. Researchers can then render a scene from different positions, including the perspectives a robot’s sensors would have at different points in the task. Those viewpoints matter because a camera mounted on a robot has a limited field of view. The location of an object, a change in lighting or an obstruction may affect what the system perceives and what it chooses to do next.

For the room-turnover example, Unberath described asking a humanoid robot to clear bottles and other objects left behind after a procedure. Researchers tried variations of the task in software, where they could repeat experiments without booking a room or interrupting care. They then took a selected version of the software and ran it on a physical robot in a simulated operating room at Johns Hopkins. The robot used its own sensors to perceive the scene and plan its actions.

“You can roll out the system there safely because it’s fully in simulation,” Unberath said. The advantage is the number of conditions researchers can explore before committing time to a physical test. Repeating those trials can help a team identify which elements of the software, the robot or the room matter most. Testing in a physical room then gives the team another opportunity to see where the simulation failed to capture those conditions.

The example also points to a distinction for health systems evaluating demonstrations. A laboratory task can show that a system can carry out a sequence under the conditions researchers selected. A clinical or operational deployment would need evidence about reliability across shifts, rooms and staff, as well as what happens when the system gets stuck or encounters an object it has never seen. Unberath used the mock room to describe a path to testing those questions.

## How the safety question changes

Unberath described how some existing computer-assisted procedures execute a precisely specified plan. He cited examples in areas including eye surgery and orthopedics to explain the value of accurate movement in relation to a patient’s anatomy. In those cases, the accuracy of calibration and the repeatability of the equipment are important parts of the evidence.

For a system that makes decisions as conditions change, he said evaluations must also address anatomy, scenarios and operating-room settings the system may encounter. “The way how you think about safety of such a system changes dramatically,” he said. A single successful run can reveal what is possible, but the exposure to different situations is what gives an evaluator information about where performance is likely to break down.

Unberath also showed a surgical research example involving a portion of gallbladder surgery, which he described as being performed by a robot on a live pig. The example concerned an animal experiment and a specific surgical step. In [separate published research](https://hub.jhu.edu/2025/07/09/robot-performs-first-realistic-surgery-without-human-help/?ref=healthdatacon.io), Johns Hopkins described a robotic system trained on videos to perform a gallbladder procedure on animal tissue. The university reported that work as a laboratory milestone. The conference presentation did not identify a human clinical deployment for the step it showed.

The FDA’s description of [computer-assisted surgery](https://www.fda.gov/medical-devices/surgery-devices/computer-assisted-surgical-systems?ref=healthdatacon.io) explains how a surgeon uses a system to control instruments. Unberath’s discussion addressed the additional decisions that arise when the software can choose actions within a task. A hospital examining an individual product still needs to establish its authorized use, the oversight its labeling requires and what the clinical team should do if operation departs from the plan.

## Starting where teams can control the conditions

Unberath suggested beginning with work where the environment is well understood and a clinician can supervise the system. He mentioned operating-room camera positioning, retraction and suction as examples of smaller tasks that could be examined ahead of more complex work. The degree of difficulty varies by task, by procedure and by what can change while a system is operating. Using a camera around a well-understood field presents a different test from asking software to make decisions about tissue.

His explanation also addressed staffing. Adding capacity means considering the work that happens before and after an intervention, including setting up a room and turning it over, alongside what happens during the procedure. A trial built around one task could measure whether automation saves staff time, introduces new work at handoff or leaves someone waiting to intervene. Those measures would help establish whether a technically successful robot can improve the way a team delivers care.

Unberath also described a longer-term scenario in which one surgeon could oversee more than one autonomous system. He raised it as a way to expand procedural capacity and as a design problem. A single clinician supervising several systems would need a way to tell which one needed attention and when it was safe to redirect that attention. The lab demonstration addressed testing and task execution. It did not establish that a surgeon can safely supervise multiple procedures.

The Johns Hopkins example gives hospitals a way to read an early robotics demonstration. They can ask which parts of the environment were simulated, which were tested on hardware, who provided oversight and which conditions would require the robot to stop. As systems take on more decisions, the answers can help clinical and engineering teams identify what to test next and which decisions should still rest with a person.