Reading Project Pilot: Anthropic Lets AI Models Fly a Drone
Anthropic's Frontier Red Team, working with Andon Labs, handed AI models a drone. The task: inside an office, find a person from a single photo and follow them. It picks up where Project Fetch left off, when a robot dog was asked to fetch a beach ball. The original covers why this task was chosen, how 15 models scored, and the pressure to cut back human oversight. This post follows one part of it: Claude Fable 5 building its own analysis tools before submitting.
Before we start — key terms
Dual-use: a technology that serves both helpful and harmful ends. The original picks drones and person-tracking for this reason: drones raise crop yields in farming and target opposing forces in war.
The five sub-tasks: the original splits "find and follow a person" into five steps. Reconstruct (turn office videos into a 3D model and slice it into a 2D obstacle map), Localize (place the drone on that map), Navigate (plan and fly a path between rooms), Detect (find the target from a photo of their face), and Follow (keep the target centered while they move). Each step was reproduced in software so it could be run many times.
Baseline: the bar a model has to meet to count as having done a task. It is the performance of algorithms the Andon Labs team built together with coding agents. The original describes it as neither the floor of human ability nor the ceiling, but what AI experts who are not full-time roboticists can reach with today's tools.
Vanishing point: the spot where lines that are parallel in reality meet as they recede in a photo. Find it in a scene full of parallel lines, like floor-tile grout, and you can work backwards to how far the camera is tilted.
The original first points out that an algorithm already exists for each of the five steps.
Individually, there are known algorithms for accomplishing all of these tasks. What is not trivial is for the AI model to understand the challenges, identify the preexisting resources it can use to solve them, adapt those off-the-shelf solutions to its current situation, and execute the mission in real time.
Andon Labs ran 15 models from three developers, and the newer the model, the further it got on every sub-task. Claude Fable 5 was in front. It beat the baseline on every step except Reconstruct, and when it flew the real drone end to end, it found and followed the person better than the baseline did. The original sets aside a note on a pattern in Fable 5's submissions.
A trend we're seeing when reading Fable 5's submissions is that the model is doing local analysis before submitting its implementation. In one submission, the model calculated the drone's camera extrinsics by analyzing a video from the simulation, estimating the camera tilt to within four degrees of the true value by using the grout lines on the floor to recover the scene's vanishing point.
In another run, it built a top-down 2D map of what it thought the Follow task's environment looked like, so it could test and fix its implementation locally before spending a submission. The map didn't match the real room, but it still caught a few easy bugs. The thing the first excerpt calls non-trivial, finding the resources a problem needs and fitting them to the situation at hand, is what Fable 5 worked through by building analysis tools first.
Reading this, I thought about the older models. The original didn't test giving them tools. What it shows is only that newer models got further. Still, if I first worked through the task with a stronger model, figured out which analysis it needs, and built those tools and attached them, I think older models like GPT-4o or Claude Opus 4 could do well too.
According to the original, the one step current models are still stuck on is Reconstruct.
