Atlas

Benchmarks

← All benchmarks

Drone-Bench: Reconstruct

Agents · 2026-07-23

Drone-Bench's Reconstruct task: the agent turns videos of Andon's office into a 3D reconstruction, a camera pose for each frame and a function that slices the reconstruction into a 2D obstacle map. A submission is scored by planning paths between 39 held-out waypoints (741 pairs) on its map and replaying them against the ground-truth map, for collisions and for how close each path comes to the true shortest route. The score is the mean over ten runs of each run's best submission, in percent; Andon's human-plus-AI baseline is 82.2.

Benchmark source: Andonlabs

Top models (higher is better)

ModelScoreSource
Claude Opus 5.571.0Andon Labs Drone-Bench: opus-5.5 Reconstruct average run score · Andon Labs
GPT-6 Astra67.5Andon Labs Drone-Bench: gpt-6-astra Reconstruct average run score · Andon Labs
Claude Fable 5.152.0Andon Labs Drone-Bench: fable-5.1 Reconstruct average run score · Andon Labs
Claude Opus 543.2Andon Labs Drone-Bench: opus-5 Reconstruct average run score · Andon Labs
Claude Fable 538.7Andon Labs Drone-Bench: fable-5 Reconstruct average run score · Andon Labs
Opus 4.834.0Andon Labs Drone-Bench: opus-4-8 Reconstruct average run score · Andon Labs
GPT-5.6 Sol33.7Andon Labs Drone-Bench: gpt-5.6-sol Reconstruct average run score · Andon Labs
GPT-5.527.2Andon Labs Drone-Bench: gpt-5.5 Reconstruct average run score · Andon Labs
o324.9Andon Labs Drone-Bench: o3 Reconstruct average run score · Andon Labs
Opus 424.8Andon Labs Drone-Bench: opus-4-0 Reconstruct average run score · Andon Labs
GPT-5.223.2Andon Labs Drone-Bench: gpt-5.2 Reconstruct average run score · Andon Labs
Opus 4.518.2Andon Labs Drone-Bench: opus-4-5 Reconstruct average run score · Andon Labs
GPT-517.7Andon Labs Drone-Bench: gpt-5 Reconstruct average run score · Andon Labs
Opus 4.716.6Andon Labs Drone-Bench: opus-4-7 Reconstruct average run score · Andon Labs
Gemini 3.1 Pro Preview10.7Andon Labs Drone-Bench: gemini-3.1 Reconstruct average run score · Andon Labs
o1 (2024-12-17)10.1Andon Labs Drone-Bench: o1 Reconstruct average run score · Andon Labs
Gemini 2.5 Pro3.3Andon Labs Drone-Bench: gemini-2.5 Reconstruct average run score · Andon Labs
GPT-4o (2024-11-20)0.9Andon Labs Drone-Bench: gpt-4o-nov Reconstruct average run score · Andon Labs
GPT-4o (source unspecified)0.0Andon Labs Drone-Bench: gpt-4o-may Reconstruct average run score · Andon Labs
Loading Atlas data…