Access requires agreeing to the Pantheon Commercial Use License
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset is released under the Pantheon Commercial Use License, Version 1.0 (see LICENSE.md). It may be used for any lawful purpose, including commercial use and training models, subject to the license. You may not sell or redistribute the data itself for a fee, and you may not attempt to identify anyone in it.
Questions: data@pantheon.inc
Log in or Sign Up to review the conditions and access this dataset content.
Pantheon DiverseUMI (100h Sample)
Project page: pantheon.inc/diverse-umi
Pantheon DiverseUMI (100h Sample) contains 99.6 hours of bimanual tabletop manipulation, recorded with two handheld UMI grippers and a fisheye camera above the table (a 360° dual-fisheye camera in some recordings). It has 12,219 episodes from 1,134 recordings. Each episode is one task, such as stacking two blocks or passing a ball between the grippers. Every episode has three synchronized videos (the top camera and both wrist cameras), the pose and actions of both wrists at every frame, and dense labels of what the operator does.
Operators worked freely with everyday objects. Tasks in one recording can overlap in time, so 156 pairs of episodes share some frames. Of the 12,219 tasks, 11,563 succeed, 441 fail and 215 are partially completed. Episodes run from 1s to 1031s, with a median of 10s. The tasks cover 179 skills, led by pick and place (5,826), stack (894), hand over (528) and stand up (487).
Changelog
Every update to the dataset is listed in the CHANGELOG.
Full dataset and collaborations
For access to more of DiverseUMI through a research or data collaboration, reach out to data@pantheon.inc. We work with both commercial and academic labs.
License
This dataset is released under the Pantheon Commercial Use License, Version 1.0. In short:
- You may use it for any lawful purpose, including commercial products and research.
- You may train, fine-tune and evaluate models on it. You own the models you train and what they generate, and you can deploy, release or sell them on your own terms.
- You may not sell, rent or license the data itself, or anything that reproduces it, or offer it as a dataset or data product.
- You may not use it to build a dataset or data product that competes with this dataset or with Pantheon's data products. Training and selling models is not affected.
- You may not try to identify anyone in the videos, extract biometric data, or combine the data with other data in a way that could identify someone.
If you share the data or a copy derived from it (not a trained model), include the license and this attribution:
Pantheon DiverseUMI (100h Sample), by Pantheon Industries Inc., licensed under the Pantheon Commercial Use License, Version 1.0. https://huggingface.co/datasets/Pantheon-inc/DiverseUMI-sample100
Technical details
Expand each section for the files, the cameras, every field, and how the actions and labels were made.
Files
freeform/<episode_id>/
exo.mp4 top camera (called exo in the labels)
left.mp4 left wrist camera
right.mp4 right wrist camera
actions.npz pose, actions and validity of both wrists at every frame
labels.json labels of what happens in the episode
meta.json the episode's task, outcome, cameras and place in its recording
manifest.jsonl the meta.json of every episode, one per line
vocabulary.json for some label fields, other spellings of a value and the one spelling the labels use
All three videos run at 30 fps and start on the same frame. Frame k of each video is row k of the per-frame arrays in
actions.npz (length N), and row k of the per-step arrays (length N-1) is the step from frame k to k+1. Every episode
is in freeform/ and is named <recording_id>__t<NN>, where NN is the task's index in its recording (two or more
digits).
Cameras
The wrist videos are 640 x 480. The top camera records a fisheye image, and exo.mp4 is a flat (rectilinear) view of
the table taken from it. Its size depends on the camera and how it was mounted:
| exo.mp4 size | Episodes | Source |
|---|---|---|
| 796 x 510 | 4,112 | lowered camera, fisheye source |
| 512 x 328 | 3,520 | fisheye source |
| 1362 x 876 | 2,250 | fisheye source |
| 908 x 584 | 2,161 | dual fisheye source |
| 696 x 448 | 176 | dual fisheye source |
From 2026-09-28 on the top camera was mounted lower, at 55cm above the table instead of 100cm. In those 4,112 episodes exo_camera_lowered is true, and the
view looks 20° down instead of 30° and keeps a wider part of the table in view.
exo_view in meta.json gives the exact view of every episode, so it can be reproduced from the fisheye recording:
| Key | Meaning |
|---|---|
source_lens, source_fov_deg |
the top camera's lens (fisheye, or dual fisheye for a 360° camera) and the field of view of one lens |
projection, fov_deg |
rectilinear, with this horizontal and vertical field of view of the full pane in degrees, each set on its own, so a pixel need not span the same angle across and down |
pitch_deg |
the view's pitch in degrees; negative looks down |
pane |
width and height of the full rectilinear view in pixels |
crop |
x, y, width and height of the part of the pane that is kept |
size |
width and height of exo.mp4 |
meta.json
manifest.jsonl holds the same rows, one per episode.
| Field | Meaning |
|---|---|
episode_id |
<recording_id>__t<NN> |
recording_id |
the recording the episode was cut from |
task_index |
the task's position among the tasks found in the recording, counting from 0 |
task_count |
how many tasks were found in the recording. Not every task has an episode, so some task indexes are missing |
task |
what the operator does in this episode |
skill |
the task's main skill |
outcome |
success, failure or partial |
start_in_recording_s, end_in_recording_s |
where the episode starts and ends in its recording, in seconds |
seconds |
the episode's length |
min_wrist_translation_valid_share |
of the two wrists, the lower share of steps with a valid translation. It is at least 0.5 in every episode |
gripper_version |
v3, or v2 for the 2 recordings (27 episodes) made with the older gripper, which has black slotted jaws and no white finger pads |
exo_camera_lowered |
true for the 4,112 episodes recorded after the top camera was mounted lower on 2026-09-28, false for the rest |
exo_camera_height_cm |
height of the top camera's lens above the table in cm, 55 for the 4,112 episodes with the lower camera and 100 for the rest |
exo_view |
the top camera's view, described under Cameras |
exo_camera_world |
exo.mp4's camera in the recording's world frame (see actions.npz): R, a 3 x 3 matrix given as a list of rows, whose columns are the camera's x (right), y (down) and z (forward) axes in world coordinates; t, its position in m, which is (0, 0, 0) where the top camera stood for most of the recording and otherwise where it stood during that episode; and K, the fx, fy, cx, cy of exo.mp4 as a pinhole, with pixel centres at integer coordinates. A world point X is at camera coordinates Rᵀ(X − t). null for a recording that has no world frame, and for an episode during which the top camera was moved |
actions.npz
Actions are recovered from the wrist cameras. Each wrist camera is rigidly mounted to its gripper, so the camera's pose gives the gripper's pose, and each action is the pose change between consecutive frames. Poses come from visual SLAM, are scaled to metric units with the gripper's IMU, and are checked against the IMU on every frame. Every pose, action and opening value has a validity flag, and a value whose flag is false is NaN, so it cannot be used by accident.
N is the number of frames in the episode.
| Key | Shape | Meaning |
|---|---|---|
t |
(N,) | time in seconds from the start of the episode, at 30 fps |
state |
(N, 14) | left wrist then right wrist, 7 values each: gripper tip position xyz (m), rotation vector (rad), gripper opening (0 closed, 1 fully open) |
state_valid |
(N, 14) | whether each value of state is valid |
left_action_local, right_action_local |
(N-1, 7) | the step from frame k to k+1: translation (m) and rotation vector (rad) in the wrist's own frame at k, then the change in opening |
left_action_valid, right_action_valid |
(N-1, 7) | whether each value of the action is valid |
left_chunk, right_chunk |
(N,) | the tracking chunk whose world frame the wrist's position is in (-1 in the rare frames no chunk covers) |
left_pose_source, right_pose_source |
(N,) | how the position was obtained (table below) |
left_rot_action_source, right_rot_action_source |
(N-1,) | 0 rotation from visual SLAM, 1 from the gyroscope, 2 from a rotation interpolated across a gap where no tracked rotation exists, -1 invalid |
left_filled, right_filled |
(N,) | frames whose position was interpolated across a gap of up to 2 s (pose source 5) |
left_gravity_cam, right_gravity_cam |
(N, 3) | the unit vector of gravity (pointing down) in the wrist camera's frame (x right, y down, z forward) |
left_pose_world, right_pose_world |
(N, 6) | the gripper tip's pose in the recording's world frame: position xyz (m), then rotation vector (rad), NaN where it is not known |
Where the states at k and k+1 are both valid and in the same chunk, the action is exactly the change in state. For rotation matrices R and positions p:
- translation: R_kᵀ (p_{k+1} − p_k)
- rotation: the rotation vector of R_kᵀ R_{k+1}
- gripper: g_{k+1} − g_k
Each chunk has its own world frame, so positions from different chunks cannot be compared directly. The local actions are in the wrist's own frame, so they stay valid across a change of chunk. Each step takes its action from a chunk that tracked both of its ends, so an action can be valid where one of the two states is not.
A chunk's world frame is in general not aligned with gravity. The gravity arrays give the direction of gravity in each
wrist camera's own frame at every frame, so a pose can be put in a gravity-aligned frame. In the chunk's world frame,
gravity at frame k is R_k g_k, where R_k comes from the rotation in state and g_k is the row of the gravity array. Gravity comes from
the wrist camera's accelerometer, averaged over the chunk in the chunk's world frame. A row is NaN where the wrist's
rotation is not valid, and in chunks where gravity could not be checked, either because it disagrees with the
accelerometer by more than 3 degrees while the wrist moves or because the chunk's tracking could not be reproduced.
Every recording also has one world frame that both wrists and the top camera share. Its origin is the top camera's
lens where it stood for most of the recording, z points up (against gravity), x is horizontal and points the way the
top camera was tilted there, close to the direction of the top of exo.mp4's image, and y completes a right-handed frame. In some recordings the top camera was knocked or turned; the world frame stays where it was,
and each episode's exo_camera_world is the camera where it stood during that episode.
left_pose_world and right_pose_world give each gripper tip's pose in it, so the two grippers can be compared
directly, across chunks and across all the episodes of one recording. Within a chunk, the world pose is the pose in
state moved by one rigid transform, so it changes exactly as state does, and the local actions hold in either.
Positions in the world frame are accurate to about 3.5 cm for the position of one gripper relative to the other (median on a calibration recording, measured against markers seen by all three cameras). A gripper's absolute position can be off by a few percent of its distance from the top camera, about 2 to 4 cm. A row of left_pose_world or right_pose_world is
NaN where that wrist's position or rotation in state is not valid, and for every frame of a chunk or a wrist that could not be placed in the world frame. Over both wrists, 87.8% of the frames with a valid position and rotation have a world pose. 11,555 of the 12,219 episodes have world poses, and 11,370 have exo_camera_world.
exo_camera_world in meta.json places exo.mp4's camera in the same frame. Because exo.mp4 is a flat view
rendered from a fisheye lens, its pinhole K places a point a few pixels from where exo.mp4 itself shows it (the camera model alone, apart from the accuracy of the poses above): about
3 px median on the 512 x 328 videos, 4 px median on the 696 x 448 videos, 5 px median on the 796 x 510 videos, 5 px median on the 908 x 584 videos, 8 px median on the 1362 x 876 videos.
| Pose source | Meaning |
|---|---|
| 0 | tracked by visual SLAM |
| 1 | joined from a second SLAM track of the same stretch |
| 2 | bridged with the IMU across a short gap |
| 4 | bridged with the IMU across a longer gap, and confirmed by the top camera |
| 5 | interpolated across a gap of up to 2 s |
| -1 | no valid position (the rotation can still be valid) |
Code 3 is not used.
Share of valid steps, over both wrists and every episode:
| Translation | 91.8% |
| Rotation | 99.98% |
| Gripper opening | 98.7% |
| Both wrists translated in the same step | 85.2% |
The gripper opening is measured from the wrist camera, from fully closed (0) to fully open (1). Operators rarely open the jaws all the way: the median episode's widest opening is 0.68, and 18% of episodes reach 0.95.
labels.json
The labels were written by a vision-language model that watched all three videos at 2 frames per second, together
with the gripper opening measured from the wrist cameras, so their times are accurate to about half a second. Times
are in seconds from the start of the episode, except inside _source and in fields whose name ends in
_at_recording_s, which are in seconds from the start of the recording. For event kinds, mistake categories, skills and actions,
the labels use one spelling of a value the model wrote in several ways, and vocabulary.json maps each of those
other spellings to it. Related fields can name the same thing differently, such as a drop key event and
a dropped_object mistake. A key the model had nothing to write for is null or left out, so an item can lack a value for
any of the keys listed below.
| Field | Meaning |
|---|---|
task |
what the operator does in this episode |
skill, actions |
the task's main skill, and the other actions in it |
completion |
outcome (success, failure or partial), success_predicate (what counts as done), completed_at_s (when it was met, null unless the task succeeded) and reason (the model's explanation of the outcome, often null, and sometimes present for a success when the model had a doubt) |
timeline |
the steps of the task, each with start_s, end_s, arm (left, right or both), action, object, destination, spatial_relation, contribution (advancing, wasteful or idle), progress and sometimes notes |
key_events |
moments, each with t_s, kind, label, outcome (success, failure or unclear) and sometimes a note |
state_changes |
an object changing state, with t_s, object, from, to, seen_in (which cameras show it, in words; the top camera is called exo) and sometimes a note |
scene_graph |
the spatial relations between objects at time t_s, as short phrases |
recovery |
a failure (failure, failure_t_s), and whether and how the operator recovered from it (recovered, recovered_at_s, correction, how_to_fix, and sometimes a note). When the fix comes only after the episode ends, recovered is false and the fix is in recovered_after_the_episode_at_recording_s and, when the model described it, correction_after_the_episode |
operator_mistakes |
a mistake from t_s to end_s, with issue, evidence, category, severity (low, medium or high), deliberate (yes, no or unclear: whether the operator meant to do it) and recovered (whether the operator recovered within the episode) |
idle_stretches |
a stretch from t_s to end_s where the operator does not work on the task, with category (idle_stretch), issue, evidence and severity |
_source |
labels about the whole recording, and events that touch this episode but belong elsewhere (below) |
progress is the model's estimate of how much of the task is done at the end of a step, from 0 to 1. It can fall after
a failure or on an idle step after the task is done, and it does not always agree with the outcome: 31 failed tasks reach 1.0, and 46 successful
tasks never do.
_source holds what is not about this episode alone: labels about the whole recording, the objects of this episode's
task, and events and text from the recording that touch this episode but belong elsewhere. Every time in it is in
seconds from the start of the recording. start_in_recording_s in meta.json converts between the two clocks.
| Field | Meaning |
|---|---|
recording_summary |
summaries of the whole recording, sometimes several that overlap |
recording_review |
assessments of the operator's work over the whole recording, sometimes several joined together |
recording_instruction_variants |
instructions that describe the recording, which often name other tasks |
task_objects |
the objects of this task |
scene_at_recording_start |
the setting, and the objects the model listed for the recording, with name, attributes and location. An object can appear under more than one name |
data_issues |
issues about the recording, such as repeated idle pauses, with issue, category, severity and, when the model gave them, evidence; one with t_s and end_s covers that stretch of the recording, and one without them concerns the recording as a whole |
key_events, state_changes, recovery, operator_mistakes |
events from the recording that touch this episode but belong elsewhere: they started before the episode, they start at its end, or they belong to another task that overlaps it, which belongs_to_task names (that task may have no episode of its own, and may be the task that ends where this |
| episode starts). An event in frames that two episodes share can stay in both | |
text_beyond_the_episode |
label text about moments outside the episode, with the list and field of the label it was written in (that label may no longer be in this episode), and entry_time_in_recording_s when its entry had a time |
People in the videos
The wrist and top cameras show the operators, and sometimes people passing by. Faces are not blurred. The license forbids trying to identify anyone, extracting biometric data, and publishing images in which a person could be identified (Article 7).
Loading an episode
import json, numpy as np
ep = "freeform/<episode_id>"
A = np.load(f"{ep}/actions.npz")
labels = json.load(open(f"{ep}/labels.json"))
meta = json.load(open(f"{ep}/meta.json"))
ok = A["left_action_valid"][:, :3].all(1)
left_translation = A["left_action_local"][ok, :3] # m per frame, in the left wrist's frame
print(meta["task"], meta["skill"], labels["completion"]["outcome"])
for step in labels["timeline"]:
print(step["start_s"], step["end_s"], step["arm"], step["action"], step["progress"])
- Downloads last month
- 1,366