Andon Labs dropped Drone-Bench - a benchmark testing autonomous drone navigation in physical space. Core task: given a photo, find and track a specific human in the real world, zero human intervention.
Axel Backlund (co-founder) confirms the obvious: fully autonomous AI-controlled drones executing search-and-follow missions in meatspace is concerning af.
DeepMind's Astra reportedly matched human baseline performance on every task variant. This isn't simulation - it's real-world spatial reasoning, visual recognition, and continuous tracking through dynamic environments.
Tech implications: multimodal models are crossing the sim-to-real gap faster than expected. Physical autonomy benchmarks like this expose how close we are to deployable embodied AI.
Axel Backlund (co-founder) confirms the obvious: fully autonomous AI-controlled drones executing search-and-follow missions in meatspace is concerning af.
DeepMind's Astra reportedly matched human baseline performance on every task variant. This isn't simulation - it's real-world spatial reasoning, visual recognition, and continuous tracking through dynamic environments.
Tech implications: multimodal models are crossing the sim-to-real gap faster than expected. Physical autonomy benchmarks like this expose how close we are to deployable embodied AI.