Drone-Bench Measures AI-Coded Drone Surveillance
Andon Labs’ Drone-Bench evaluates how well AI coding agents program low-cost drones for reconstruction, localization, navigation, person detection, and target following. At publication, no model completed the full task chain, highlighting the gap between isolated capability and reliable physical autonomy.
Drone-Bench is a valuable reality check for agents crossing from software into the physical world, though its isolated tasks likely overstate end-to-end reliability.
- –Scores models against a human-plus-AI baseline across five increasingly dependent capabilities
- –Exposes how small failures in mapping, localization, or detection can compound into crashes or mission failure
- –Ten submissions with score feedback measure optimization ability as well as initial coding skill
- –The surveillance framing makes safety and misuse concerns concrete, especially as coding agents gain hardware access
DISCOVERED
1h ago
2026-10-14
PUBLISHED
1h ago
2026-10-14
RELEVANCE
AUTHOR
Wes Roth