You receive trajectories, not splats.
A policy has to output actions video doesn't contain — so we deliver the calibrated, action-labeled episode your pipeline already ingests. Below is the exact shape of what lands in your bucket.
Three formats, your ingest spec
Reinforcement Learning Datasets — the TFDS-based episodic format behind Open X-Embodiment. Native for many VLA pipelines.
The Hugging Face community standard. Parquet + video, easy to load, replay and fine-tune with the LeRobot stack.
The underlying hierarchical store — arbitrary tensors per timestep, ideal when you have a custom loader.
The volumetric rig and 4DGS are the means we use to extract precise, calibrated 3D action and hand-pose tracks. No buyer ingests a splat — you ingest the trajectory.
# One episode — RLDS / LeRobot compatible, HDF5 on disk
episode/
├── observation/
│ ├── image.cam_high float [T, 720, 1280, 3] # multi-view RGB
│ ├── image.cam_side float [T, 720, 1280, 3]
│ ├── image.cam_wrist float [T, 480, 640, 3]
│ ├── depth.cam_high float [T, 720, 1280] # optional RGB-D
│ └── state float [T, 8] # joint pos (7) + gripper (1)
├── action float [T, 7] # 6-DoF Δpose + gripper ← trained target
├── hand_pose float [T, 21, 3] # per-frame 3D finger joints (dexterous)
├── force_torque float [T, 6] # Fx Fy Fz Tx Ty Tz (contact-rich option)
├── language_instruction str # "pick the mug and place it on the rack"
├── is_terminal bool [T]
├── reward float [T] # sparse success signal
└── meta/
├── camera_intrinsics float [K, 3, 3]
├── camera_extrinsics float [K, 4, 4]
├── embodiment str # target action space / retarget notes
├── consent_id str # per-contributor consent artifact ref
├── environment str # scene / lighting / geography tag
└── sync float [T] # genlock-validated timestamps (ms)“Trainable out of the box” has a testable definition.
Every delivery ships against agreed acceptance criteria — because poor sync or bad calibration silently corrupts observation–action pairs, and you shouldn't find out at training time.
Delivered trajectories replay to their logged end-state in your (or our) sim/robot above an agreed threshold.
Per-camera intrinsics/extrinsics held within a stated error budget — reported per delivery, not hand-waved.
Genlock-validated timestamps; each observation–action pair is time-aligned within a documented tolerance.
6-DoF Δpose + gripper (and hand pose where dexterous) verified against the recorded demonstration, not interpolated.
Environment / object / demographic spread reported against the agreed scoping matrix.
Every engagement ships a Datasheet for Datasets.
Collection process, sensor list, calibration procedure, known biases and intended use — the documentation a Head of Data expects. A reference episode and full datasheet are available to inspect under NDA during scoping, so you validate the format before you commit.