Glossary

The hub keeps its own words — the technical one stays on the page, and a dotted underline shows this sentence on hover, focus or tap. Here they all are, grouped the way the Quickstart meets them, each with the page it is first explained on.

Robots

MJCF: MuJoCo's robot description file (XML); the hub simulates and renders from it.Seen on a robot's page
URDF: ROS's robot description file (XML); the hub converts it to MJCF when you add a robot.Seen on the sign-up page
floating base: The robot's body moves through the world, like a legged robot — unlike a fixed base, bolted down like an arm.Seen on a robot's page
fixed base: The robot is bolted down, like an arm; only its joints move.Seen on an arm's page
kp, kd (gains): Each joint servo's stiffness (kp) and damping (kd): how hard it pulls toward its target and how much it resists motion.
interface: The robot's own declaration of what it is to a policy: joint order, base, default pose, gains, what an action means, the control rate.Seen on a robot's page
fingerprint: A short code computed from the interface (or a file); the same code means nothing changed.Seen on a robot's page
embodiment: Which robot body a policy was trained for, recorded as the interface fingerprint it was stamped with.Seen on a robot's Policies tab
match (embodiment check): This policy's embodiment fingerprint equals the robot's current interface: it was trained for exactly this body.Seen on a robot's Policies tab
mismatch (embodiment check): This policy was trained against another interface; a connected robot refuses it.Seen on a robot's Policies tab
unstamped (embodiment check): This policy was stamped before embodiment fingerprints existed, so only its IO contract is checked.Seen on a robot's Policies tab
undeclared (embodiment check): The robot declares no interface, so there is nothing to compare the policy against.Seen on a robot's Policies tab

Policies and training

policy: The trained program that turns what the robot senses into the motor commands it sends — what a run produces and a robot runs.Seen on the sign-up page
run: One training job: a template, a robot, a few settings and a budget; it ends in a model or a failure, with its log and checkpoints kept.Seen on the sign-up page
checkpoint: A snapshot of the policy saved partway through training, so an earlier point can be tested or deployed.Seen on the sign-up page
template: A ready-made training recipe — which algorithm, which simulator; you pick one and set a few numbers.
reflex: A small, fast policy that runs on the robot itself at a fixed rate: balance, walking, holding a pose.Seen on a model's page
skill: A larger policy that takes camera images and a task and plans a chunk of actions; it is served from a computer and the robot follows.Seen on the train form
planner: A policy class that decomposes a task into steps for other policies; it never drives a joint itself.Seen on a model's page
control class: Whether a model is a reflex or a skill — it decides where it may run and what the robot checks before running it.Seen on a model's page
class undeclared: This model does not say whether it is a reflex or a skill, so nothing can say where it may run; stamp the policy with lucen model stamp --control-class.Seen on a model's page
observation: Everything a policy gets as input each step — joint positions, velocities, gravity direction, the last action — as one vector.
contract: What a policy expects in and sends out: the observation layout, the action size and scale, the control rate.Seen on a model's page
ONNX: The one-file format a trained reflex policy is exported to; what a robot loads.
executor: Where a job runs: on a runner you start on your own computer, or on the hub's cloud GPUs.
runner: A small program you start on your own computer; it picks up your queued jobs and runs them there.Seen on the sign-up page
claim: A runner taking a queued job for itself, so no other runner does.
cloud (Modal): Rented cloud GPUs the hub runs jobs on, where a deployment has them configured.

Sim tests

sim gate (sim tests): Sim tests: running a policy in the simulator through a fixed set of tests and scoring each PASS or FAIL, before it goes near hardware.Seen on Sim tests
rollout: One sim-test job: a policy on a robot through a set of test cells, with a scorecard and a video per cell.Seen on Sim tests
battery: The set of test cells a rollout runs: smoke is a few quick ones, full is the whole table.Seen on the sign-up page
seed: The random starting point of a test; the same seed replays the same test, and more seeds make a score more trustworthy.

Serving a skill

endpoint: A skill model served from a GPU over a network connection, so a robot can ask it for its next actions.Seen on My robots
endpoint capable: The robot's adapter can hold and stop on its own, so this robot may consume a hub-served endpoint.Seen on My robots

Datasets

episode: One recorded attempt in a dataset: the frames from the start to the end of a task.Seen on a dataset's page
frame: One time step of a dataset — the robot's state, the action and the camera images at that instant.Seen on a dataset's page
LeRobotDataset: Hugging Face's layout for robot datasets: parquet tables, mp4 videos and a meta folder — what the hub reads.Seen on a dataset's page
state (dataset feature): Per frame, what the robot measured: one number per joint, in the order the dataset names.Seen on a dataset's page
action (dataset feature): Per frame, what the robot was told to do: one target per joint, in the order the dataset names.Seen on a dataset's page

Connected robots and consent

device: A robot you connected: its computer runs lucen-device, which pairs with the hub and refuses anything not approved.
heartbeat: The signal a connected robot sends every few seconds; a minute without one and it counts as offline.Seen on My robots
trust mode: What a connected robot will run: only policies a person approved (the default), or any of its registered contracts.Seen on My robots
approval request: A request to run one policy on one robot for a set number of minutes; a signed-in person approves or denies it.
approval token: A short-lived signed ticket the robot verifies itself — how it knows a person approved this exact policy.

Agents and keys

API key: A secret string that lets a terminal, a script or an agent act as you; the browser never needs one.Seen on API keys
scope: What a key may do: read looks, write changes your repos and files, train also starts jobs that can cost money.Seen on API keys
scoped key: An API key made with one scope — read, write or train — so whatever holds it acts as you only that far.Seen on the sign-up page
MCP: The protocol coding agents (Claude Code, Cursor) use to call tools; the hub ships a server for it.Seen on API keys
campaign: A series of training rounds toward one goal on one robot, run by the agent.
plan: The agent breaking one task on one robot into steps, each handoff approved by a person.

Files and ids

digest: A checksum of a file's bytes; the hub recomputes it, so what you download is what was uploaded.Seen on a model's page
ULID: The long id in mono (run_01M4…): unique and sortable by time — copy it, never type it.

Accounts

handle: A person's username or an organization's slug — the first part of every /{owner}/{repo} URL; the two share one namespace.Seen on Organizations
organization: A handle a lab or a team shares: robots, models, datasets and runs published under one name, by every member.Seen on Organizations
owner (of an organization): An organization member who may add and remove members, change roles and edit the profile; every organization keeps at least one.Seen on Organizations
member (of an organization): A person in an organization who may publish under its handle; an owner adds them by their hub handle.Seen on Organizations