Repository navigation
The arm on LeRobot 0.7.0
LeRobot 0.7.0 replaced feetech-servo-sdk, the library under the SO-101's bus, with rustypot, one
SerialMotorsBus for every Feetech and Dynamixel servo. quackd[lerobot] installs LeRobot 0.7.0 now,
below 0.7.1, with rustypot below 1.14, and every LeRobot name the arm relies on was read again at the
v0.7.0 tag, with the bus below LeRobot's Python at rustypot's v1.13.1.
What the arm says changed with the bus. A failed torque call names every joint it could not switch,
so connect attempt 1 of 3 failed on elbow_flex (id 3) and gripper (id 6): ... sends you to both
cables. A release over motors that do not answer gets the time LeRobot 0.7.0 takes, a panic in
rustypot is met as a failed bus call, and a second program on the arm's port is refused as it opens
it, on Linux and macOS too. quackd robot release, and a --by-hand run that opened the gripper,
close it before they let go of the arm.
lerobot:mock runs the real backend's own code over an arm in memory, so a rehearsal on it refuses,
hands off and closes in the arm's own words.
Industrial arms on ROS 2
quackd[rosarm] installs the tenth adapter, quackd-rosarm, for a Universal Robots UR5e, a KUKA LBR
iisy 3 R760 or LBR iiwa 7 R800, or a Yaskawa MotoMini, each behind a ROS 2 driver, its maker's for
three of them and the LBR-Stack for the iiwa, and rosbridge. A move is one goal to every joint, inside
the travel the arm's own description gives and at most a tenth of each joint's rated speed, and every
move asks a person first
(ADR-0078).
quackd never powers, enables, un-brakes or un-stops an arm, parks none and drives no tool.
rosarm:mock rehearses it over a fake rosbridge and a fake of each arm's driver, and rosarm:ws
refuses every arm until somebody names an owner and a date for its first run. A workflow in CI,
rosarm contract, runs the transport against rosbridge_server on Jazzy and Humble with a stand-in for
the arm's driver
(the industrial arm's page).
A Universal Robots UR7e joins the described bodies beside the UR5e.
A body's load, a key held down and a heartbeat on 3.11
A body can say what rides on its back or deck, carries_kg, and a task can ask for it
(ADR-0079).
A key held down, or an Enter read late by a key thread that a busy machine held up, no longer answers
yes to a question that drops an arm. A run on Python 3.11 ends though its heartbeat's cancel was lost.
The policy server
The policy server reads LeRobot 0.7.0. It loads nothing where the kernels package is installed,
refuses a pinned model whose config would run a kernel from the Hub and a GR00T whose VLM step names
no model, and fetches a GR00T base pinned at a tag at the commit its config.json was read at.
What changes for you
quackd[lerobot]installs LeRobot 0.7.0 and rustypot, which Python 3.14 on Linux builds from
source.- A second program on the arm's port is refused on Linux and macOS too, so close whatever holds it
before a run. quackd robot release, and a--by-handrun that opened the gripper, close it before they let go
of the arm.quackd_lerobot.real.load_policy()is gone.list-adapters --jsonanddoctor --jsonsay"extra": nullfor a row no package installs, where
they said"built-in".- Pillow 13 is quackd's own floor.
- The policy server refuses a SmolVLA, pi05, pi0 or X-VLA checkpoint that LeRobot 0.7.0 saved, for
now, and loads nothing wherekernelsis installed, souv pip uninstall kernelsin its
environment if it is there.
What has not run
No arm has run quackd on LeRobot 0.7.0. The SO-101 last ran it on 2026-11-06, under quackd 0.31.3
and LeRobot 0.6.1, and
its checklist
and the note's Known limitations list what the arm owes. No controller and no arm has met rosarm,
and no described body has been compared with its robot.
Eleven distributions, one version between them: quackd, quackd-microduck, quackd-lerobot,
quackd-rosbridge, quackd-open-duck, quackd-xlerobot, quackd-alohamini, quackd-toddlerbot,
quackd-huebot, quackd-moss and quackd-rosarm. The full note is in
CHANGELOG.md.