mmahdavian/isaac_sim_augmented_realobjs

UR5e cup-to-dishrack obstacle demonstrations

This training release contains 42 accepted simulated demonstrations, 46,306 synchronized robot samples and 126 RGB videos, captured at 30 Hz. There are 26 obstacle-removal demonstrations and 16 detours. The robot is a UR5e with a Robotiq 2F-140 gripper.

The dataset follows the LeRobot v2.1 folder layout of /work/hamid/real2sim/real2sim_ep_03_B. At the user's request, it retains this batch's original 30 Hz, seven-dimensional measured-state and commanded-action conventions. It does not adopt the reference's 15 Hz, 14-dimensional commanded-state or next-frame-action convention. The videos are copied byte-for-byte without resampling or re-encoding.

Layout

data/chunk-000/episode_000000.parquet
videos/chunk-000/observation.images.wall/episode_000000.mp4
videos/chunk-000/observation.images.wrist/episode_000000.mp4
videos/chunk-000/observation.images.body_forward/episode_000000.mp4
meta/info.json
meta/episodes.jsonl
meta/episodes_stats.jsonl
meta/tasks.jsonl
meta/source_mapping.jsonl
meta/simulation/episode_000000/{scene_spec,training_validation,report}.json
robot_data/episode_000000.hdf5
assets/ur5e/norm_stats.json

Every episode has a Parquet file, three MP4s and its original lossless synchronized.hdf5 recording (renamed under robot_data/). meta/source_mapping.jsonl maps the new contiguous episode indices to original source IDs and scenarios. The earlier round_01/ video release is preserved; its IDs are original source IDs, so consult the mapping rather than equating filenames across layouts.

State, actions and camera alignment

• observation.state: measured [shoulder_pan, shoulder_lift, elbow, wrist_1, wrist_2, wrist_3, gripper_open_fraction].
• action: absolute commanded targets in that same order. Arm angles are radians; gripper openness is 1 - native/255 (1 open, 0 closed).
• At time t, all three images and measured state are sampled before the action applied over [t, t + 1/30). Actions are neither next-frame measured joints nor pre-normalized deltas.
• observation.velocity: six measured joint velocities, radians/second.
• observation.tcp_pose: robot-base position in metres and quaternion [x,y,z,qw,qx,qy,qz].
• timestamp is relative episode time. sim_timestamp, action_end_timestamp and camera_timestamps preserve simulator timing in float64.
• source_episode, variation, measured/commanded native gripper values and finger-joint angle provide additional provenance.
• The raw HDF5 files also preserve all 12 articulation-joint positions, velocity feedforward, camera poses, phases, joint names and capture attributes. Original state/action precision is retained there; training vectors use float32.

Camera names are physical roles: wall is the fixed countertop view, wrist the moving gripper camera, and body_forward the robot-base-mounted view. They correspond to the reference's countertop, left and right. All images are 640 by 480 RGB.

Splits and training

Training episodes are indices 0–36 (37 episodes). Validation episodes are 37–41 (5 episodes), corresponding to original recordings 7, 14, 21, 28 and 35. Every augmentation inherits its original recording's split. Source episode 5 failed simulation validation and is excluded; this is the validated 42-of-43 subset.

meta/episodes_stats.jsonl holds per-episode statistics. meta/stats.json describes all episodes; use meta/train_stats.json for training-only feature statistics. assets/ur5e/norm_stats.json contains training-only OpenPI normalization for a 32-action horizon: arm targets are converted to deltas relative to the first measured state in each chunk, and the gripper remains absolute. The stored Parquet actions remain absolute. At 30 Hz, a 32-action chunk spans about 1.067 seconds.

The tested reader is LeRobot v2.1 at commit 0cf864870cf29f4738d3ade893e6fd13fbd7cdb5, with datasets==3.6.0 and PyAV. For that reader:

from huggingface_hub import snapshot_download
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

repo_id = "mmahdavian/isaac_sim_augmented_realobjs"
root = snapshot_download(repo_id, repo_type="dataset",
    allow_patterns=["data/**", "videos/**", "meta/**", "assets/**"])
train = LeRobotDataset(repo_id, root=root, episodes=list(range(37)),
    delta_timestamps={"action": [i / 30 for i in range(32)]}, video_backend="pyav")
validation = LeRobotDataset(repo_id, root=root, episodes=list(range(37, 42)),
    delta_timestamps={"action": [i / 30 for i in range(32)]}, video_backend="pyav")

Do not apply the reference dataset's 14-D adapter or commanded-state interpretation without adapting it to these documented fields. This release supplies data, not trained model weights.

Validation and provenance

meta/source_validation.json records full decoding of all 126 original videos, frame counts and timestamps, synchronized numeric checks, and SHA-256 hashes. meta/loader_validation.json records actual LeRobot decoding and action-chunk checks. checksums.json inventories the packaged files. All exported states/actions are checked against the raw recordings, including episode-end padding; train/validation source IDs are disjoint.

The inherited simulator checks cover placement, contact, obstacle clearance/detour and camera identity. Physical masses/friction are estimates, and robot self-collision was disabled. This is UR5e simulation data with B1K-inspired timing, not an official BEHAVIOR-1K task dataset. Depth and fictional robot joints are not synthesized.

The scanned-object library derives from choijoshua16/ur5e-scan-objects, with this project's cup, rack, marker and pen-holder assets. This release contains rendered data and simulation metadata, not source meshes. No new license grant is asserted for third-party assets. Earlier accepted episodes retain the object policies active when they were generated.