Exea LabsSheet A-00
Arrival
Opening survey data 0 / 9
Survey → Twin → Simulation → Dataset
One scan of a kitchen becomes a physics twin. A robot arm and two walking people run in it, and every successful run is written as a GR00T-format dataset. This page draws from that export. The room is an authored sample, and no real kitchen has been scanned yet.
The 3D view paused because the browser dropped the graphics context. It resumes when the browser returns it.
Sheet A-01Survey
Plan index click to select
This browser could not start WebGL, so this page shows the frames the robot recorded for each task.
04 · Dataset Sheet A-04
Each run is written in the LeRobot v2.1 layout that NVIDIA GR00T reads, with an eight-number state and action per frame plus the ego and wrist cameras. Six steps produce it.
Step 1
An iPhone LiDAR scan (Apple RoomPlan) captures the walls, counters, sink, stove and fridge. The twin on this page comes from the authored sample, since no real kitchen has been scanned in this repository.
Step 2
The compiler normalises the scan into fixtures with sizes, positions and colours, plus the drawers, doors and hinges they carry. Objects come from a small catalogue with mass and friction.
Step 3
MuJoCo steps the room at 500 Hz. People are kinematic capsule walkers on seeded waypoint loops, and every step records their closest approach to the robot.
Step 4
A scripted inverse-kinematics expert performs each task. A run counts only if the physics says it worked, such as the drawer being open or the bowl being in the sink.
Step 5
The export writes frames, states, actions and two 128 px cameras as LeRobot v2.1 in the GR00T layout.
Step 6
A loader check reads the dataset back the way a trainer would, windows the actions, decodes the videos and reports any problems.
Step 1scan
Frames written
across episodes from task families
Control rate
Hz
this page plays them back resampled to fps
State and action
/
seven joint angles plus gripper opening, per frame
Cameras
at px
ego_view and wrist, shown live in the stage
05 · Evaluate Sheet A-05
The evaluation harness ran a small batch with the scripted expert while people walked through the room. The figures below come from the export.
Revision schedule
Three things are not done. GR00T has not been fine-tuned on this dataset, so no policy has learned from it. No real kitchen has been scanned and run through the pipeline, and the room on this page is authored. The hosted demo that takes many rooms is not built. Below is every other open task, grouped by what it needs. A diamond is filled only when its pull request is merged on main.
06 · Close Sheet A-06
The twin, physics and dataset on this page were produced by the pipeline in this repository. A GR00T fine-tune and a scan of a real kitchen are both still missing, so this is a working pipeline without a result.