Astra flies a drone and beats the human baseline on Drone-Bench
Given a drone and video of an unfamiliar office, Astra builds its own 3D map, plans a route and flies, beating the human baseline on every Drone-Bench task in at least one attempt, though only 2.8% of runs succeed end to end.
GPT-6 Astra just crossed a line for me. Give it a drone, videos of an unfamiliar office, and the goal of finding and following a specific person. It can build its own navigable 3D model, plan the route, fly the drone, and has beaten the human baseline on every Drone-Bench task in at least one attempt. Yes, reliability is still bad. Only 2.8% of runs succeed end to end. But that almost makes the result more interesting. The underlying capabilities are already there. One general model is figuring out perception, mapping, planning, tool use and physical action on a task it wasn’t specifically built for. If AGI means general intelligence that can adapt to genuinely new problems across domains, I honestly don’t know what else we’re waiting for.
This post quotes another post: x.com/andonlabs/status/2098103320208712049