GPT-6 Astra can navigate drones to find and track people
The GPT-6 Astra model was tested for its ability to create an autonomous human surveillance system using a regular, low-cost drone. In the tests, the model had to solve tasks of navigating an office, recognizing a specific person from a photograph, and subsequently tracking them.
This is stated in the Drone-Bench study by Andon Labs.
Drone-Bench was created to evaluate how well modern artificial intelligence models are able to write software for autonomous surveillance of physical environments using inexpensive drones.
The demonstration scenario requires the drone to independently navigate an office, find a specified person, and follow them. To achieve this, the entire process was divided into five tasks: creating a 3D model of the office, determining the drone’s location, navigating between rooms, recognizing the person, and tracking them.
At the recognition stage, the system receives a photograph of a specific person’s face and must find precisely that person in the video stream from the drone’s camera. Upon detecting the target, the system controls the drone to follow the person, keep a specified distance, and not lose them from the frame.
Among the models that Andon Labs checked in Drone-Bench is GPT-6 Astra. The researchers also tested previous models from OpenAI, Claude, Gemini, and Fable.
Andon Labs noted that, starting with Claude Opus 4.8, the average result of new models already exceeds the baseline system created by the researchers in the separate tasks of human recognition and tracking.
«It is a bit creepy to watch,» – noted the authors of Drone-Bench.
At the same time, the study does not confirm that GPT-6 Astra is already capable of completely autonomously executing the entire scenario from start to finish. Each of the five tasks is tested separately, and for the next stage, the models are provided with the correct results of the previous one.
According to Andon Labs, currently no tested model has surpassed the baseline system in creating a 3D model of the room. Due to the accumulation of errors, the success rate for completing the entire scenario from building a map to autonomous human pursuit remains at 0%.
The authors of Drone-Bench expect that the very next generation of models will be able to pass all five stages in their best attempt. Andon Labs emphasizes that with the growth of AI autonomy, such systems become easier to use for unintended purposes.
Politics Editor
Read more similar news:
Comments:
comments powered by Disqus
