Dongkkka/Task_OrangeCoffee_Mobile_Sim50_Visual8x_HeadCam_LeRobotV30

Task OrangeCoffee Mobile — Sim50 Visual8x HeadCam (LeRobot v3.0)

Synthetic mobile-manipulation dataset for the ROBOTIS FFW-SG2 Rev1, formatted as a LeRobot v3.0 dataset.

Task

Move the right orange coffee can with the right hand and the left orange coffee can with the left hand to the matching cabinet on the right.

The policy controls both 7-DoF arms, both grippers, and lateral mobile-base motion (linear_y). The head and lift joints are not part of the state/action vectors.

Composition

• 450 episodes and 726,480 frames at 15 Hz (about 13 h 27 min of data)
• 50 unique Isaac Sim trajectories (80,720 frames)
• 8 visual-domain-randomized rerenders per trajectory (400 episodes and 645,760 frames)
• 50 original renders + 400 rerenders = 450 total episodes
• One task and one RGB head-camera view
• Train split only (0:450); no validation or test split

The 400 augmented episodes preserve the source robot/can/cabinet poses, state/action arrays, episode lengths, and timing. Only appearance and camera rendering are randomized. Therefore this dataset contains 450 rendered episodes but only 50 unique state/action trajectories.

Visual randomization

The rerenders randomize:

• dome-light intensity and color
• background-wall RGB color
• coupled camera focal scale (0.95–1.05)
• camera roll, pitch, and yaw (up to ±1 degree each)
• visual yaw of each coffee-can label (−180 to +180 degrees)

Robot, coffee-can, and cabinet geometry is protected during visual replay.

Modalities

Feature	Shape / format
observation.images.rgb.cam_left_head	RGB video, 3×376×672, H.264/yuv420p
observation.state	float32, 17 dimensions
action	float32, 17 dimensions

State and action order:

1. arm_l_joint1 … arm_l_joint7
2. gripper_l_joint1
3. arm_r_joint1 … arm_r_joint7
4. gripper_r_joint1
5. linear_y

linear_y is the commanded/measured lateral mobile-base component. Other base axes are excluded from this task-specific policy contract.

Origin and generation

The 50 source trajectories were generated in Isaac Sim using the Cyclo Lab Task000002TrajectoryGenerationEnvCfg environment and then passed through a strict success/audit selection step. They were assembled from three valid generation shards containing 7, 27, and 16 successful episodes. The selected HDF5 source was task_000002_orange_transport_physical_curated50_commanded_hold_v12.hdf5.

In this provenance, “physical trajectory” means physics-aware simulation generation. This dataset does not contain real-robot demonstrations.

Runtime contract:

• physics: 120 Hz
• control: 15 Hz
• camera: 15 Hz
• swerve controller: 30 Hz

Detailed per-episode generation and randomization records are included in:

• meta/isaac_action_replay_provenance.json
• meta/source_replay_manifest.json
• meta/physical_runtime_rate_provenance.json

Known limitations

• Only 50 independent state/action trajectories are present; the other 400 episodes are visual rerenders of those trajectories.
• The dataset is synthetic and contains no real-robot samples.
• It covers one instruction, one task, and one head-camera viewpoint.
• It has no held-out evaluation split.

Model association

This dataset was used to train Dongkkka/Task_OrangeCoffee_Mobile_450augmented_GR00T_Intern.