k6sol

quackd v0.16.0 is out! The SO-101's pilot can now hand the arm to a learned policy for one short segment at a time. A trained ACT drove the arm's simulator twin, and GPT-6 Sol from OpenAI used it for two manipulate segments on that twin. For example, the checkpoint natsuki0000/act-so101-bluecap ran two segments on the twin through serve-mcp. The next day gpt-6-sol handed the same ACT two manipulate calls under --goal and --policy-url; the run failed because the scene held a red cube and a pen instead of the blue cap the policy had learned. The policy itself lives in quackd policy serve (ACT, SmolVLA or pi05), the arm side is only an HTTP client, and every segment is confirm-gated, step-capped and clipped to the arm's travel. In plain terms: the pilot still decides what the next short piece of work is, then hands the arm to a policy that was trained on it. The policy runs in its own process so the bus stays light, motion is rate-limited, and anything that looks unsafe aborts and holds. You can rehearse the whole thing on the twin first. None of it has driven the real arm yet. github.com/rokbenko/quackd/r… @huggingface @LeRobotHF @OpenAIDevs