Researchers at King’s College London and Toshiba Europe have connected a robot arm to an edge computer through a complete 5G network, replacing repeated camera-image uploads with compact descriptions of the workspace. In fruit-packing and cube-stacking trials, the largest gains across three methods were a 52.6% reduction in task completion time and a 45-percentage-point increase in success probability against a system that periodically sent RGB images. The unreviewed September 14 study moves a technique previously tested mainly in simulation onto physical hardware.
Edge-assisted physical AI divides a control loop between a machine in the room and a nearby server. A camera shows the server what surrounds the robot, an AI model chooses the next action, and a command returns to the arm. Sending a fresh image at each step consumes bandwidth and forces the server to interpret many pixels that have no bearing on the task. High image compression can also delay the loop because a small onboard computer must spend time encoding every frame.
The team’s goal-oriented communication system sends the facts needed for the next action. The robot detects objects in colour and depth images, then converts the scene into 3D bounding boxes or a scene graph, a list of relationships such as an apple standing on a table. These records cross OpenAirInterface, an open-source implementation of a 5G network, to an Nvidia Jetson AGX Orin edge server. A language model selects a motion; the 3D variants can reconstruct the scene in a digital twin to check for collisions before a command returns to the arm.
To test the full loop, the researchers used a PiPER arm and inserted random failures such as dropped objects, missed grasps and collisions. Fruit packing and three-cube stacking were each repeated 20 times with the image baseline and with each of the three semantic formats. The radio used a 40-megahertz channel across 0.5 metre, so the trial measured software and protocol delays under a short, controlled link rather than a busy factory network.
The communication measurements show where time was saved. On the 40-megahertz 5G link, a 150.06-kilobyte RGB frame at the highest PNG compression setting took 597.86 milliseconds from encoding through decoding; encoding alone accounted for 543.66 milliseconds. A 0.0383-kilobyte 2D scene graph completed the path in 17.84 milliseconds. Structured data also gave the edge server a narrower reasoning task than interpreting a full image, contributing to the reported gains in completion time and recovery from the inserted faults.
A robot network could therefore transmit the state required for its next decision. The trade-off between representations remains unresolved: 2D graphs carry spatial relationships cheaply, while 3D graphs add depth and collision checks at extra processing cost. The preprint covers two tabletop tasks, one half-metre link and 20 runs per framework. Longer links, shared radio traffic, varied lighting and multiple robots will show whether the timing and recovery gains survive in warehouses and factories.
Related coverage
Sources
- Shutong Chen and colleagues, “Goal-Oriented Communications for Physical AI: Design and Testbed,” arXiv, September 14, 2026. Abstract, authorship, submission date and principal results.
- Full paper. System sequence, hardware, network configuration, experiments, results and limitations.
- Project page. Testbed overview and supporting material supplied by the authors.
- Lead image: National Institute of Allergy and Infectious Diseases, Wikimedia Commons, public domain. The photograph is illustrative and does not show the reported experiment.
