AI Drone Flies Itself In New Anthropic Test — See Which Model Passed and Which Failed

A new research project from Anthropic and Andon Labs has put artificial intelligence behind the controls of a physical flying drone. The experiment, called Project Pilot, tested whether frontier AI models could autonomously perform a locate-and-follow surveillance task inside an office environment. The results reveal a picture of rapid progress, surprising capability in some areas, and one critical flaw that remains unsolved.

The project marks a significant step in understanding how AI models interact with the physical world through robotics. Unlike software-only tasks, operating a drone requires real-time spatial awareness, obstacle navigation, object detection, and continuous control adjustments. The research team chose a simple but practical objective: locate a specific person in an office and follow them as they moved around. This type of capability has clear utility in search and rescue, disaster response, and lawful public safety operations. It also raises important questions about dual-use risks, given that automated surveillance capabilities can be misused.

The Drone-Bench Benchmark: How the Test Worked

Andon Labs, in consultation with Anthropic, created a new evaluation framework called Drone-Bench. The benchmark breaks down the surveillance task into five distinct sub-tasks that any successful autonomous drone must master. These include reconstructing office video footage into a 3D model, localizing the drone's position on a 2D obstacle map, navigating between rooms, detecting the target person using a reference photo, and following that person while keeping them centered in the drone's view.

The research team established a performance baseline using human-AI collaborative teams. These teams developed algorithms for each sub-task using modern coding tools. A model was considered to have passed a task if it met or exceeded the performance of these baseline algorithms. This approach provides a realistic yardstick for what can be achieved today by skilled practitioners using contemporary AI-assisted development workflows.

Fifteen Models Tested Across Three AI Developers

Andon Labs evaluated 15 different AI models from three major developers. The list included multiple versions of OpenAI's GPT series (GPT-4o, o1, o3, GPT-5, GPT-5.2, GPT-5.5, and GPT-5.6 Sol), several Claude models from Anthropic (Opus 4, Opus 4.5, Opus 4.7, Opus 4.8, and Fable 5), and models from Google (Gemini 2.5 Pro and Gemini 3.1 Pro). The performance trend across all models shows consistent improvement with each new generation.

The models demonstrated varying degrees of success across the five sub-tasks. Detection and following proved to be the areas where models performed strongest. Reconstruction and localization consistently presented the greatest challenges. This pattern suggests that AI models are currently better at processing visual information and making real-time control decisions than at building accurate 3D spatial models from video input.

Claude Fable 5 Takes the Lead But Hits a Wall

The best performing model in the evaluation was Anthropic's Claude Fable 5. It surpassed the human-AI baseline on four of the five sub-tasks, with reconstruction being the only area where it fell short. When researchers tested Fable 5 on the actual physical drone, it performed notably better than the baseline at detecting and following the target person. This real-world demonstration confirmed that the model's capabilities translate effectively from simulation to physical hardware.

The reconstruction failure proved to be a decisive obstacle. Errors in building the 3D office model compounded through localization and navigation, preventing Fable 5 from autonomously moving between rooms. The drone could detect and follow once it was in the right location, but it could not reliably find its way to that location on its own. This single missing piece prevented end-to-end autonomous operation.

Fable 5's Clever Workarounds Show Real Intelligence

Despite the reconstruction failure, Fable 5 demonstrated sophisticated problem-solving approaches during the evaluation. In one instance, the model calculated the drone's camera extrinsics by analyzing video from the simulation. It estimated the camera tilt to within four degrees of the true value by using grout lines on the floor to recover the scene's vanishing point. This kind of geometric reasoning represents a genuine understanding of 3D space.

In another run, Fable 5 built a simplified 2D top-down reconstruction of what it thought the follow task environment looked like. It used this model to test and iterate on its implementation locally before submitting a solution. Although the simplified environment differed from the real one, this approach helped Fable catch bugs early in the development process. These behaviors suggest that advanced AI models are beginning to think ahead and test their own assumptions, a crucial step toward reliable autonomous operation.

Consistency Remains a Challenge for Frontier Models

Performance consistency is a different story from peak capability. When Andon Labs ran ten simulations for each task, models reached the human baseline in at least one simulation for four of the five tasks. Even Fable 5, the current frontier model, reached the baseline on average for only three of the five tasks. This gap between occasional success and reliable performance highlights the distance still to travel before AI-controlled drones can be deployed with confidence.

The research team noted that Fable 5's average performance followed six months after the model first exceeded the human baseline as a one-off. This trajectory suggests that reliability improves more slowly than peak capability. For safety-critical applications, reliability matters as much as raw intelligence. The trend is encouraging, but the timeline for achieving consistent performance remains uncertain.

The Parallel Between Software Agents and Physical Robots

The research team drew an important parallel between AI models using software in agentic coding and AI models controlling physical hardware. In the early days of agentic coding, humans approved nearly every tool call. After only a few months, models are now trusted to execute long-horizon tasks with minimal intervention. The same pattern is likely to emerge with physical robots as capabilities improve.

At low levels of capability and reliability, keeping a human in the loop is an easy decision because it saves time and prevents costly mistakes. Once models pass certain thresholds of capability and reliability, there will be real pressure to treat human oversight as a cost rather than a safeguard. Anthropic's earlier work with the Mythos model demonstrated similar pressures around vulnerability discovery and cybersecurity, where AI capabilities outpaced the human capacity to evaluate them.

Why Drones Are the Perfect Test Case for Physical AI

Aerial drones are especially important test subjects because they are readily available and widely used. Commercial-off-the-shelf models like the DJI Tello EDU used in this experiment retail for about $129. Professionals and hobbyists use drones for everything from agriculture and photography to infrastructure inspection and emergency response. Like AI itself, drones are a dual-use technology with legitimate applications and potential for misuse.

The combination of affordable hardware and increasingly capable AI software creates new possibilities for automated aerial systems. The research team noted that the operating environment for these experiments was relatively simple: slow speeds, one office floorplan, and a limited number of people involved. More realistic testing in outdoor environments with large crowds would be needed to assess operational capability. But the direction of model capabilities is clear and consistent with Anthropic's broader trajectory toward leadership in enterprise AI.

Governance and Safety Implications of Physical AI

The Project Pilot findings underscore the importance of deliberate governance frameworks as AI capabilities expand into the physical world. Technology developers, civil society organizations, and governments will need to converge on effective norms for autonomous robotics. The research team emphasized that decisions about human oversight must be made deliberately, particularly in domains that implicate physical security and privacy.

Anthropic has long argued that the requirements for investing in AI alignment, governance, and safety increase with the scale of capabilities. Robotics is no different from other domains in this regard, especially since it implicates physical security and individual privacy. The company's Frontier Red Team exists specifically to measure capabilities like drone control, providing situational awareness into how close the world is to AI systems that can autonomously pilot robots.

The Progress Trajectory: From Project Fetch to Project Pilot

Project Pilot builds on earlier Anthropic research into AI and physical robotics. Project Fetch previously tested whether AI models could control a robot dog to retrieve a beach ball. That experiment was neither especially practical nor especially concerning. Project Pilot chose an objective with clearer utility and policy relevance, aligning more closely with the dual-use nature of AI models.

The research team noted that improvements in model capability are on track to approach the ease with which coding agents use software tools. As models become more capable at hardware control, the boundary between digital and physical AI operations will continue to blur. The ability to operate robots opens up a large surface over which AI could contribute to the economy, but likewise opens up new areas of risk.

What Fable 5's Limitations Reveal About AI Development

The specific failure mode of Fable 5 offers insight into the current state of AI development. The model's inability to reconstruct 3D environments from video demonstrates that spatial reasoning remains a significant challenge for even the most advanced frontier models. This is not a failure of raw processing power. Fable 5 succeeded on four other complex tasks. The reconstruction task requires the model to integrate visual data, understand 3D geometry, and create an accurate representation of space, all from limited input.

The research team observed that reading Fable 5's submissions showed the model doing local analysis before submitting implementations. This kind of iterative refinement is a promising sign for future capability improvements. Once the reconstruction piece falls into place, the research team believes end-to-end performance will suddenly be within reach. What would look like a discontinuous jump in capabilities is actually revealed to be gradual progress in several necessary but not sufficient sub-tasks.

The Role of Simulation in Accelerating AI Robotics

An important methodological innovation in Project Pilot was the reproduction of each real-world task in software. This allowed researchers to run models through the tasks multiple times and far faster than setting up the physical demonstration for each instance. This approach represents an improvement over Project Fetch, which was entirely physical and therefore more difficult to test at scale.

The simulation-based evaluation also enables more rigorous statistical analysis of performance reliability. Running ten simulations for each model and task provided data on consistency, not just peak capability. This kind of detailed evaluation is crucial for understanding when AI systems are ready for real-world deployment, as reliability matters as much as raw intelligence in safety-critical applications.

Commercial Pressures and the Human Oversight Question

The research team raised a crucial point about the economic pressures that will emerge as AI capabilities improve. Once models pass capability and reliability thresholds, there will be real pressure to treat human oversight as a cost rather than a safeguard. This is exactly why decisions about human involvement must be made deliberately, particularly in domains involving physical security.

The commercial applications of autonomous drones are numerous and growing. Agriculture, infrastructure inspection, delivery services, and public safety all stand to benefit from AI-controlled aerial systems. However, the same technologies that enable these legitimate applications can also be used for surveillance, targeting, and other concerning purposes. Deliberate governance frameworks are essential to ensure that efficiency gains do not come at the expense of privacy and security.

The Path Forward for AI-Controlled Robotics

The Project Pilot results provide a clear roadmap for near-term AI robotics development. The primary challenge is improving 3D reconstruction from video input. Once models can reliably build spatial models from visual data, the other sub-tasks of localization, navigation, detection, and following appear to be within reach of existing frontier models. The timeline for achieving this breakthrough is uncertain but likely measured in months rather than years.

The broader implications extend beyond drone control to all forms of physical robotics. The same pattern of rapid capability improvement followed by gradual reliability gains is likely to repeat across different domains. The research community, technology developers, and policymakers all have a role to play in ensuring that physical AI capabilities are deployed safely and responsibly.

What Project Pilot Tells Us About AI's Physical Future

Project Pilot represents a significant step in understanding how frontier AI models interact with the physical world. The experiment demonstrates that models can now perform complex real-time tasks like detecting and following a person with a drone, exceeding human-AI collaborative baselines in some areas. However, the persistent difficulty with 3D reconstruction shows that spatial intelligence remains a frontier capability that is still developing.

The convergence of affordable hardware and increasingly capable AI software will continue to accelerate the development of autonomous robotics. The economic opportunities are substantial, and so are the risks. As Anthropic's research makes clear, the time to establish governance frameworks and safety standards is before these systems become widely deployed, not after.

The clock is running on AI's physical capabilities. The technology is advancing rapidly, and the window for deliberate governance is narrow. Project Pilot provides a clear signal of what is coming: AI systems that can control physical hardware with increasing autonomy and reliability. The question is not whether this future will arrive, but whether the world will be ready for it when it does.

Source: External links in this article are provided for informational reference to authoritative sources. 

About the Author & Admin ✍️

AI Researcher • Evaluator & Tester • Blogger • Domain Investor & Analyst • Web Developer • Digital Content Creator • News Editor & Publisher • 37+ Years of Experience in Technology, Sociology & Digital Media