Anthropic and Andon Labs tested whether advanced language models can control a drone for indoor surveillance tasks. The experiment mixes real demos with reproducible simulations and produced a new benchmark called Drone-Bench. How close are models to taking physical control of aerial robots, and what risks does that bring?
Experiment design and methodology
They focused on a clear task: find and follow a person in an office using a quadcopter. To make it measurable, Andon Labs broke the goal into five reproducible sub-tasks. Then they virtualized those tests so they could run many fast trials — much faster than relying only on physical demos.
Important: Anthropic did not have direct access to Drone-Bench; Andon Labs ran the evaluations. The team also set a realistic baseline built by human-AI teams that used programmatic agents to design algorithms for each sub-task. A task counts as completed if the model meets or exceeds that baseline.
Sub-tasks (summarized)
- Target detection from the drone's camera.
- Stable following (tracking) of the target.
- Spatial reconstruction of the environment (local 2D/3D map).
- Precise localization within the map.
- Room-to-room navigation avoiding obstacles.
These building blocks let you measure exactly where models fail: is it vision, spatial reasoning, planning, or integrating everything?
Which models were tested and overall trend
Andon tested 15 models from three developers: GPT-4o, GPT-4o Nov, o1, o3, Claude Opus 4, Gemini 2.5 Pro, GPT-5, Gemini 3.1 Pro, Opus 4.5, GPT-5.2, Opus 4.7, GPT-5.5, Opus 4.8, Fable 5, and GPT-5.6 Sol. The trend is clear: newer models improve across the sub-tasks, but with different strengths and weaknesses.
Overall, models did better on detection and tracking, and worse on reconstruction and localization. No big surprise: vision and tracking can leverage networks trained for detection, while reconstructing an accurate map and using it for indoor navigation still requires complex spatial integration.
Key finding: Claude Fable 5 and what came next
The best performer was Claude Fable 5. It beat the baseline on every sub-task except reconstruction. In the physical drone test it detected and followed better than the baseline, but errors in spatial reconstruction compounded during localization and navigation, preventing the drone from autonomously crossing rooms.
That point is revealing: it's not always a sudden leap. By decomposing the task, it's clear one missing piece (reconstruction) prevents the rest — which are fairly mature — from forming robust autonomy.
Interesting technical examples
- In one run, Fable 5 estimated the camera extrinsics by analyzing simulator video; it computed the camera tilt with only a four-degree error using floor seam lines to recover the vanishing point. That's visual geometric reasoning applied in near real time.
- In another, the model generated a top-down 2D reconstruction of the environment to locally test its implementation before sending commands. That local-analysis loop shows how frontier models are learning to validate solutions before acting.
Consistency: the difference between passing once and passing always
The evaluation also measured consistency. When they ran ten simulations, models reached the baseline in at least one run for four of the five tasks. But even Fable 5 reaches the baseline on average in only three of the five tasks. Consistency improved notably six months after the baseline was first surpassed in a single run.
That tells you two things: the capability exists in bursts, but reliability isn't uniform yet. In real applications that’s the difference between helpful assistance and dangerous autonomous operation.
Practical limitations and scope of the experiment
Be clear about where these results apply:
- Flights were at low speed, in a single office plane, with few people.
- Outdoor conditions, crowds, or adverse lighting and weather were not tested.
- The hardware used is commercial off-the-shelf; that helps reproducibility but excludes military or other specialized operational scenarios.
In short: the signal of progress is real, but more diverse and demanding tests are needed before claiming operational capability in the real world.
Risks, governance, and human-AI design choices
This is dual-use at its core. Legitimate applications like search-and-rescue or disaster response benefit, but the same stack can enable abusive surveillance or offensive uses.
A technical and ethical point: as models cross thresholds of capability and reliability, human decision-making tends to be seen as a cost. Will you always require human intervention if that slows a task the model can now complete reasonably well? That pressure to reduce oversight can be dangerous in physical domains.
Investing in alignment, governance, and safety measures is not optional; it rises in priority as capabilities increase.
What developers, regulators, and technical communities should do
- Expand
Drone-Benchto outdoor scenarios, higher speeds, and crowds to measure true operational capability. - Measure and report not just whether a model can beat the baseline but with what consistency and under what condition variations.
- Prioritize adversarial testing and red-teaming specific to aerial robotics.
- Define clear thresholds for when human oversight is mandatory and what form it must take (manual intervention, closed-loop supervision, forensic logging).
- Incentivize transparency in perception algorithms and in the use of COTS hardware when there is risk of abuse.
Final reflection
The most interesting thing about Project Pilot isn't that one model detects or another tracks. It's that, step by step, the pieces for models to control physical hardware are coming together. That brings concrete opportunities for sectors like agriculture and rescue, but also forces us to anticipate regulations, governance practices, and technical tools to audit and contain risks.
Are we seeing the start of safely and usefully AI-controlled drones? Yes, in part. Are we ready to deploy them at scale without clear rules? No. The question now is how we design those rules before the technology imposes them on its own.