quackd-lerobot

The LeRobot SO-101 arm adapter for quackd, and the arm's simulator. Install it through quackd rather than directly:

uv pip install "quackd[lerobot]"       # the mock, and the real arm through LeRobot on Python 3.12+
uv pip install "quackd[lerobot-sim]"   # the simulator: MuJoCo and no LeRobot, on Python 3.11+
uv pip install "quackd[lerobot-vla]"   # the policy server: LeRobot, torch and transformers, on Python 3.12

One SO-101 has run this adapter, on 2026-09-15 and again on 2026-09-23. Every LeRobot name it relies on is read from upstream source at a pinned commit and listed in upstream_api.py.

lerobot:mujoco is the real backend's own code over a physics model of the SO-101, the maker's model from TheRobotStudio's SO-ARM100, fetched at a pinned commit on first use and never shipped. quackd robot twin makes one of a registered arm, and quackd preflight rehearses task files on it. Nothing has compared it against an arm. Every name it reads from the model is in sim/upstream_api.py.

quackd policy serve runs a learned policy for the arm, an ACT, a SmolVLA or a pi05 checkpoint, in a process of its own and never beside the serial bus, and --policy-url hands it the arm's pick and manipulate. No policy has driven an arm through it yet: docs/guides/policies.md.

What it does, what it refuses and why: docs/adapters/lerobot/README.md.