finde159/DataEngine_NT

Franka cube stacking — generated trajectories with IDM pseudo-actions

7,232 action-labelled manipulation trajectories, 2.40 M frames, 66.7 hours at 10 Hz. Every frame in this dataset was generated by a video model. No robot was moved to produce it.

A Cosmos-Predict2 LoRA imagines a Franka FR3 stacking three cubes; a fine-tuned inverse dynamics model watches the imagined video and recovers the actions that would produce it. No human selected any sample at any stage — trajectories were kept or discarded by an automatic filter.

Conforms to the Physical Data Engine pde/training-dataset@v1 output contract, in LeRobot V3 layout. Read § Deviations before training on it.

Contents

episodes	7,232
frames	2,401,024 (66.7 h)
rate	exactly 10 Hz, 332 frames per episode
tasks	6 — every ordering of three cubes
views	third_person_0, third_person_1, wrist — 320×180 RGB
size	14 GB

Each episode is a four-subtask sequence: pick A → stack A on B → pick C → stack C on AB.

action.eef_delta_metric        float32[7]  [dx_m, dy_m, dz_m, drx_rad, dry_rad, drz_rad, gripper_closed]
action.eef_delta_metric_valid  bool[7]     per-dimension validity

Translation is metres per 0.1 s step in the robot_base frame for the fr3_hand body. Gripper is 0 = open, 1 = closed. The final row of every episode carries an all-false validity mask — no action is fabricated for a frame with no successor.

Loading

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("finde159/DataEngine_NT")

Or directly:

import pandas as pd
df = pd.read_parquet("hf://datasets/finde159/DataEngine_NT/data/chunk-000/file-000.parquet")

Videos are concatenated per file; meta/episodes/ gives each episode's window as from_timestamp / to_timestamp within its video file.

How it was made

1. Scene variants. A held-out camera frame is edited to move the cubes. Only the left view is edited by hand; the right view is derived by fitting each cube as a rigid 5 cm cuboid and projecting it through the fixed stereo geometry, so both views stay geometrically consistent.

2. Generation. For each of 20 layouts × 6 cube orderings, a tree of clips is generated 3 seeds wide and 4 subtasks deep — 120 clips yielding 81 complete trajectories per tree. Each clip is conditioned on the previous clip's last frame.

3. Labelling. An inverse dynamics model fine-tuned on 16 real teleoperated demonstrations reads each trajectory and emits actions. It sees two frames 16 apart and predicts a 16-step horizon, run at stride 1, so each frame is covered by up to 16 overlapping predictions.

4. Filtering. Three rules on the gripper signal, per subtask quarter. Thresholds derived from real held-out demonstrations, never tuned on generated video.

   • C1 exactly one gripper transition per quarter, in the direction the subtask implies
   • C2 that transition falls at 0.30–0.93 of its quarter
   • C3 at least 50% of the quarter's frames are unanimous across the 16-window ensemble

9,720 trajectories were generated; 7,232 (74.4%) passed and are published here. Pass rate declines along the chain — 0.99 / 0.92 / 0.86 / 0.84 for subtasks 1–4 — because each clip is conditioned on the previous one's final frame, so error accumulates.

Models: Cosmos-Predict2 LoRA and the fine-tuned IDM.

Deviations from the contract

Both were raised with the receiving team and waived. They are recorded here, and in meta/pde_output_contract.json, so nobody has to rediscover them.

The action is realised motion, not the issued command. The contract asks for desired_relative_target / command_at_t. These values are the IDM's estimate of the motion between observations. On the source rig the realised motion tracks the recorded human command at about 90% (1.824 m of path against 2.002 m commanded), so the two are close but not the same quantity. For generated video no human command exists at all. The real command channel is preserved in the native archives.

No camera calibration exists. The contract requires intrinsics, extrinsics and a frame graph. This rig has none, and these views are synthetic anyway — no physical camera was ever calibrated. meta/calibration.json states this rather than inventing values.

Things that will surprise you

The source recordings are mislabelled 60 Hz. All 20 demonstrations measure 41.68 Hz (sd 0.35), and every downstream fps label inherited the error — the intermediate datasets are tagged 15 Hz while carrying 10.42 Hz of content. This dataset corrects it: the video is retimed to real time before resampling, so 10 Hz here means 10 Hz, and 332 frames really is 33.2 seconds.

Rotation is identically zero, and that is intentional. The teleoperator never rotated the controller while clutched — the angular command is zero on every recorded row of every demo. These are not unknown dimensions filled with zero; they are a recorded hold.

The gripper label leads visible finger motion by roughly 20 frames in the source data (one frame of convention plus physical actuation lag). A visual open/closed detector will disagree with these labels systematically; the labels match the robot's control input.

fr3_hand vs the TCP. The source records the gripper TCP, offset [0, 0, 0.1034] m from the flange (measured by forward kinematics). Converting a TCP delta to a hand delta needs d_hand = d_tcp − (R(t+1) − R(t)) @ offset; because the rotation is constant here that term is exactly zero, so the translation deltas are already fr3_hand deltas. The export asserts that precondition per episode rather than assuming it.

Limitations

• One task, one scene. Three 5 cm cubes on a white table, a fixed camera rig, an FR3 arm.
• The filter checks label coherence, not task success. All three rules read the IDM's own output. A clip where the arm closes on empty air with correct timing passes. Hand inspection is the only check on physical plausibility.
• Generated video degrades late in an episode. Roughly 16% of fourth subtasks fail the structure check; artefacts accumulate along the conditioning chain.
• No held-out split. The video model was trained on 19 demonstrations with no holdout, so do not evaluate against those demonstrations expecting generalisation.

Licence

Derived from NVIDIA Cosmos-Predict2 weights under the NVIDIA Open Model Licence, and from NVIDIA's GR00T-Dreams IDM. The licence position on model outputs differs from redistributing weights; this dataset is gated for that reason. Attribution to NVIDIA is required.