jonathan

No better way to celebrate Memorial Day than with robots! 🦾🇺🇸 GPUs were spinning thru the night and now I’ve got ACT policy chaining at 100% success after ~50 trials. Watch this: one policy places the orange duck, the next removes it. Clean handoff every time. TLDR: Instead of one giant combined policy, I train small atomic ones (place specific color, remove it, etc.) and stitch them together. The technique: capture an image from the overhead cam, feed it to a multimodal LLM (OpenAI) alongside a few labeled workspace examples + a good prompt. The AI then selects the right policy or flags any error states. Has anyone tried this atomic + LLM-selector approach with VLA models? I'm curious how the performance compares vs using a larger generalized model.