repos

Search text (matches slug and description).

Filter by owner handle (username or org slug).

Filter by tag; repeatable (AND semantics).

Sort key, always descending (newest first).

UI-API: owner/slug of a robot repo — only model repos whose model_meta.robot names it (the policies built for that robot), as part of the query, never a filter over a page. A robot repo the caller cannot see yields an empty page, exactly like an unknown owner: this filter is not an existence oracle.

K2: only robot repos whose card's CURRENT interface_fingerprint is this — the robot a stamped policy was built for, found from the embodiment_fingerprint in its manifest. Part of the query, like every filter here.

UIV2-thumbs — tags adds facets to the page: how many of the caller's visible repos carry each tag, counted in the query (visibility is a WHERE, never a post-filter, so a private repo never adds to a count a stranger sees). kind and owner scope the counted set; q, tag, visibility and robot do NOT, so a filter rail's counts hold still while you click (the rule of listRuns?facets=true). Omit it and nothing is counted.

Page size.

Opaque cursor from the previous page's next_cursor.

Page of repos.

Request failed validation.

Owner handle; defaults to the caller's username.

Repo created.

Missing or invalid credentials.

Authenticated but not allowed (visibility, membership or scope).

State conflict (duplicate handle/slug, wrong repo kind, terminal job, ...).

Request failed validation.

Repo owner handle (username or org slug).

Repo slug.

The repo.

Resource not found (or hidden from the caller).

Repo owner handle (username or org slug).

Repo slug.

Updated repo.

Missing or invalid credentials.

Authenticated but not allowed (visibility, membership or scope).

Resource not found (or hidden from the caller).

Request failed validation.

Repo owner handle (username or org slug).

Repo slug.

Repo deleted.

Missing or invalid credentials.

Authenticated but not allowed (visibility, membership or scope).

Resource not found (or hidden from the caller).

UIV2-thumbs — present only when listRepos?facets=tags.

Owner handle; defaults to the caller's username.

{owner}/{slug} — the canonical address.

Present only when kind is dataset and meta has been set.

Present only when kind is model and meta has been set.

Present only when kind is robot and a card has been set.

UIV2-thumbs — on a robot repo, how many model repos name it as model_meta.robot and are visible to the caller, counted in one grouped query per page under the same visibility WHERE as listRepos (so it equals the number of repos listRepos?kind=model&robot={full_name} pages through, and never counts a private policy for a stranger). Null on dataset and model repos.

Every visible repo in the counted set (the "All" row).

One bucket per tag carried by at least one repo of the set, count descending then tag ascending; at most 100 (a rail does not print more). A repo counts once per tag.

Free-form robot type label (e.g. laika, so101).

Policy family (e.g. act, smolvla).

Base model reference when finetuned.

Training run (run_...) that produced this model, if trained on the hub.

Which layer the policy belongs to (M15). Null means undeclared — such a model cannot become an endpoint until it says.

UI-API: owner/slug of the robot repo the policy was built for. Null when none was recorded — or when that robot repo is not visible to the caller, so a public policy never names a private robot.

UI-API: the embodiment-interface fingerprint recorded in the policy's manifest (M17b stamps it): the interface_fingerprint of the robot card the policy was trained against. Null for a policy whose manifest records none — every policy published before M17b.

UI-API: embodiment_fingerprint against the recorded robot's CURRENT interface_fingerprint, computed on read. Null when no robot is recorded or it is not visible to the caller.

Map of joint name → limits.

Repo-relative path of the MJCF file inside this repo.

Markdown notes rendered on the robot page.

The embodiment interface (M17): the robot-side declaration every training, sim-gate and deploy path reads. Null on a card that only describes the hardware.

sha256 of the canonical interface JSON — the index space a policy or a dataset is built for. Null when interface is.

UI2: the last validation report stored with the card, as the publisher sent it. Null when none was sent.

UIV2-thumbs: the robot's still, one per theme, as repo files — fetch them through the repo's raw route (presigned on request). Null when the publisher rendered none (no GL on its machine, or --no-thumbnail).

Files whose blob is verified; equal to count means every digest is proven.

The tag; listRepos?tag={value} selects exactly these repos.

Joint names in policy index order. Hashed; never reorder silently.

Joint name → position (rad or m) the robot starts from and offsets are relative to.

Name of the MJCF keyframe holding the default pose, if there is one.

Floating base only.

Observation terms the REAL robot can supply (joint_pos, joint_vel, base_ang_vel, projected_gravity, foot_contact, ee_pos, …). A policy's observations must be a subset.

Bodies or sites that touch the ground (legged robots).

mujoco.__version__ for a mujoco record; null for static.

R1: the hosted check (rchk_…) this report came from, set by publishRobotCheck. The card was published exactly as that check drafted and hashed it, however long after checked_at the person pressed Publish — so a page reads the record as current rather than dating it against the card's updated_at. Still the publisher's claim, like every other field here.

sha256 over what the stills were rendered from (the renderer's version, MuJoCo's, every file the MJCF reads and the pose the card declares). lucen robot publish renders again only when it changes, so re-publishing an unchanged robot uploads nothing on any machine. Null for stills the publisher did not render (a photo).

position_offset is relative to default_pose.

Overrides scale for the joints it names.

False for a real tool the model does not contain.

Which image slot of a VLA this camera feeds.

The morphology hint generic task templates key on.

Command name → [lower, upper].

Path into the card body the check is about, when there is one.

Numbers behind the verdict (drift, penetration, counts), per check.