Coding agents given a simulator and $20 now write robot programs that beat planners built by experts. I looked inside 112 of those programs: no new algorithms, just known methods tuned to the scene, the robot, and the agent’s own habits.
A flock with no leader, three local rules, and one knob that flips a crowd from chaos to a single moving body — and what that says about planning for teams of robots.
A cost-sensitive neighborhood aggregation method for GNNs that routes neighbors by similarity — and what its failures reveal about two fundamentally different kinds of heterophily.
A simple change in perspective — planning paths for items instead of robots — leads to provably optimal warehouse rearrangement and up to 2x faster completion times.
A plain-language look at “Motions in Microseconds” (Thomason, Kingston and Kavraki, ICRA 2024), which cuts robot motion planning from milliseconds to microseconds with vectorized sampling-based planning.