We’re introducing Drone-Bench, a benchmark where AI agents code drones to complete a simple autonomous surveillance task. Drone-Bench is independent but based on Project Pilot, our work with Anthropic.
I know zero about this, but is that rigid reconstruct-localize-navigate-detect-follow flow really the way to do it? It feels overstructured.
It feels to me like if you asked me to walk into a room and track a person, I wouldn't try to map the whole area first. I'd just search for the person and dynamically avoid obstacles that appeared while heading for them. Sure, I'd build a mental map of the space, but that would be a continuous background thing; I'd start moving way before I had a full map.
The alternative is to let the AIs take actions like "go forward 3m", "rotate 25 degrees", etc at every iteration. But we found that 1) this makes the movement way too slow to follow anyone and 2) models are really bad at this.
2jbash
What about "initiate 3m of forward movement with collision avoidance", and the next "turn" starts instantly?
Hey everyone!
We’re introducing Drone-Bench, a benchmark where AI agents code drones to complete a simple autonomous surveillance task. Drone-Bench is independent but based on Project Pilot, our work with Anthropic.
I would make a condensed version here for LW, but visuals are much better on web: https://andonlabs.com/evals/drone-bench
Very curious on your feedback!