Notes on robotics, planning, neuro-symbolic AI, and the occasional thought on research life.
A shared yardstick for robotic assembly, built on LEGO Duplo: 400 tasks across four difficulty tiers, an open-vocabulary Assembly-by-Disassembly baseline, and a finding — structured planning beats a modern vision-language-action model at every tier, and the gap widens with complexity.
A short tour of TAMPanda, now public: a MuJoCo wrapper for the Franka Panda that takes you from a symbolic plan to a physically validated, collision-free execution — so you can work on the algorithm, not the plumbing.
An opinion I formed the hard way — for long-horizon robot tasks, the durable scaffold is symbolic and legible. Learning is best poured into that structure, not wrapped around it.
