On November 6, 2026, NVIDIA presented research results built on its publicly available tools for wireless network experiments and simulation, including Aerial and Sionna. About a year after Aerial's release, the company says more than 8,000 research and academic institutions worldwide use the related platform. One of the most notable results comes from robotics research by teams including King's College London. Compared with a conventional approach that periodically sends camera images, the new method reduced task completion time by up to 52.6%. However, this result was obtained using existing 5G equipment, and it measures not only communication but also inference and robot motion. To evaluate research aimed at future 6G, it is necessary to distinguish between improvements in communication speed itself and efficiency gains in the overall task that uses the communication.
Sending object positions and relationships instead of images
What the robotics research changed is how information captured by the camera is conveyed over the network. In the conventional approach, RGB images were compressed in PNG format and sent periodically, and a vision-language model on the receiving side analyzed each image to decide the next action. The proposed method instead extracts the information needed for the task on the robot side in advance and transmits only that, an approach called goal-oriented communication. Rather than conveying the whole image, it selects and sends what the robot needs to decide its next move.
Specifically, it uses 3D bounding boxes that indicate the position and size of objects, and "scene graphs" that express spatial relationships between objects. Instead of sending raw pixel data, it conveys in structured form what is where and how each object is positioned relative to the others. The research team tested both 2D and 3D scene graphs, and the method using 3D scene graphs also transmitted object bounding boxes. Its distinguishing feature is that it changes the transmitted information itself rather than simply raising image compression.
The experiments used a PiPER robot arm and an RGB-D camera, connected to the network through a 5G modem. On the Jetson AGX Orin, the team used the Sionna Research Kit to build a setup handling the base station, core network, and edge processing. The 5G network used OpenAirInterface, with a radio link distance of 0.5 m, a carrier frequency of 3.3 GHz, and a bandwidth of 40 MHz. The experiment therefore does not recreate an environment where many robots move over a wide area.
The team targeted two tasks, packing fruit into a box and stacking cubes, and carried out each task 20 times for each communication method. The tests covered grasp failures and drops, as well as placement in the wrong position, incorrect stacking order, and collisions. The results have been published as a preprint on arXiv (2609.15895).
The reported reduction of up to 52.6% is a comparison of task completion time against the conventional method of periodically sending images. The measurement includes information extraction, communication, and inference, and, depending on the method, digital-twin verification of motions, with time recorded until the robot's final motion finished. The team also reports improved task success rates, but the same time savings were not obtained for every task or communication method.
Reducing transmitted data and inference load at the same time
The proposed method adds a step in which the robot extracts the necessary information, which introduces new processing time. Even so, total task time fell because the amount of transmitted data dropped substantially and the receiver's inference became lighter. In exchange for spending time on preprocessing, it reduces the time needed for the communication and decision-making that follow.
According to the paper, a PNG image with the highest compression setting was about 150 KB, whereas the extracted information was under 1 KB. When images are compressed heavily, the compression itself takes time. The difference between the two methods therefore lies not only in the amount of data transferred but also in the time spent compressing images versus extracting information. Evaluating overall efficiency requires comparing these processing times as well.
The AI model used for inference also changes with the information sent. The method that sends only 3D bounding boxes used a large language model (LLM), while the methods using 2D or 3D scene graphs used a small language model (SLM). If relationships between objects are made explicit beforehand, the receiver has less work to do in analyzing images to understand them. The study stands out for redesigning not just the communication data format but also the AI model that processes the information.
Digital twins, which verify robot motions in a virtual space before execution, were also built into some of the methods. They were used in the method sending 3D bounding boxes and in the one using 3D scene graphs. Because the information extraction method, inference model, and motion verification setup differ from one method to another, the reduction in task time cannot be attributed to the communication method alone.
Interesting results also emerged on the relationship between radio bandwidth and latency. When the transmitted information was small enough, widening the bandwidth from 20 MHz to 40 MHz had only a limited effect on reducing latency. As the time for data transfer itself shrinks, fixed overheads that arise regardless of data volume, such as communication protocol processing, become relatively larger. This means that widening bandwidth alone may not sufficiently improve overall task latency.
Better neural receiver performance and lower base station power
SoftBank built a Transformer-based "neural receiver" into Aerial's 5G physical layer and tested whether it improved reception performance. A neural receiver uses a neural network to recover transmitted information from received wireless signals. While the robotics research changed the format of the information sent, this work changes how received radio signals are processed.
The experiment used a commercial 5G smartphone in a shielded room. It used the n78 band, a 100 MHz bandwidth, and one uplink spatial layer, and was evaluated under conditions with interference. The team prepared 15 power-control settings and, for each method, took three measurements of about one minute at each setting. The neural receiver's average uplink physical-layer throughput was reportedly higher than the conventional method at every setting.
The largest improvement was confirmed under conditions where the calibrated signal-to-interference-plus-noise ratio (SINR) was about −0.6 dB. The neural receiver reached 47.9 Mbit/s against 29.9 Mbit/s for the conventional method, and SoftBank reports an improvement of up to 60.5% based on unrounded measurements. However, this figure represents the bit rate of data successfully decoded on the receiving side. It is not a measurement of effective speed when transferring files in a smartphone application.
The same smartphone was used for training and evaluation, with identical device placement. Evaluation did, however, use a wider range of power settings than training. Going forward, it will be necessary to confirm whether the improvement holds when the number of uplink spatial layers is increased to two, and under different devices and propagation environments.
NVIDIA also introduced research in which ETH Zurich tuned a receiver using wireless measurement data collected on site. After tuning the receiver on the Aerial Testbed, the block error rate was reduced by more than two-thirds compared with the pretrained model chosen as the baseline. Block error rate indicates the proportion of received data blocks that could not be correctly recovered. While SoftBank's experiment showed higher throughput, the ETH Zurich study focuses on reducing reception errors.
Meanwhile, research at Leeds tackled reducing the power consumption of the radio access network (RAN). According to NVIDIA, the system predicts communication demand up to 10 minutes ahead, confirms that neighboring base stations can maintain users' connections, and then puts some base stations to sleep. If needed, it brings base stations back online to respond to changes in demand. In verification with a 5G test facility using DGX Spark and a digital twin, RAN power consumption fell by 45.6% while maintaining connectivity, compared with a baseline in which base stations always run at full power. However, this figure applies to the RAN; it does not mean that the power consumption of whole facilities, including GPUs, or of the entire communication network fell by the same proportion.
Research results toward 6G and remaining verification challenges
CTTC worked on replacing the radio channel calculations that the CPU conventionally handled in 5G-LENA, a module of the network simulator ns-3, with GPU processing using CUDA and WebGPU. The goal is to speed up calculations for wireless environments with many cells, users, and large antenna arrays. What is shortened here is not the time actual wireless communication takes, but the simulation time for designing and evaluating communication networks.
However, even if simulation speed improves, results are hard to use for comparing network designs if their reproducibility and reliability suffer. CTTC says it performed calibration based on 3GPP TR 38.901 and compared network performance with conventional CPU results, verifying that statistical behavior is preserved. What matters is that it did not simply speed up calculations on GPUs but also confirmed consistency with existing calculation methods.
The results NVIDIA presented are each evaluated with different metrics. The robotics study's figure of up to 52.6% is the reduction in task completion time, including communication, inference, and motion; SoftBank's up to 60.5% is the improvement in uplink physical-layer throughput. Leeds's 45.6% is the reduction in RAN power consumption, and the CTTC study targets faster simulation calculations. All relate to improving wireless network efficiency, but because what was measured and the experimental conditions differ, the numbers cannot be compared directly to judge which technology is better.
What NVIDIA announced is progress in work that makes use of the research platforms it has released. The figure of more than 8,000 research and academic institutions shows usage of the related platform as tallied by the company; it does not indicate the success rate of research results or the rate of commercial adoption. Moreover, many of the results presented were obtained using 5G equipment and experimental environments, and they do not guarantee the performance of future commercial 6G networks as they stand.
Still, it is significant that researchers can use a shared open platform to compare and verify results while changing communication methods, AI models, and network configurations. The key challenge going forward is whether improvements in task time and reception quality hold up when more robots are added, when the wireless environment becomes more complex, and when processing load rises. If such research advances, it could lead to more practical network design that considers not just peak line speed but also the overall processing time and power consumption of the systems that use the network.