Team-RAMEN/ramen-g1-ikea-assembly-ramen-ori-insert-table-leg

RAMEN G1 IKEA Assembly — RAMEN-Ori training data: insert_table_leg

One assembly skill, insert_table_leg (BitRobot task "insert table leg to table base"), cut out of BitRobot/G1_WBT_Dex1_Building-Children-Table by Team RAMEN for the IROS 2026 Humanoid IKEA Assembly Challenge (Unitree G1 + Dex1-1 assembling a children's table). It is the training data of our RAMEN-Ori policy: the all-skills model and the insert_table_leg model.

• 1,664 episodes, 482,461 frames at 30 fps (about 4.5 hours). Each episode is one segment of a source episode.
• Segments (chosen with our curation labels, task_index 0): all optimal segments.
• Size: about 1,333 GB, most of it in the pre-rendered image caches (below).

Contents

path	content
data/, meta/, videos/	A LeRobot v3.0 dataset with the source's features, unchanged (below)
frame_cache_v2_tars/overlay/	Head-left camera frames with the YOLO-OBB detections drawn on them, in 10 variants per frame (the original and 9 augmented). This is what RAMEN-Ori trains on
frame_cache_v2_tars/token/	The same head-left frames in the same 10 variants without the overlay and without rotation or shift (for a model that takes the detections as separate tokens)
frame_cache_v2_tars/common/	The two wrist cameras in the same 10 variants (no overlay)
frame_cache_tars/	The decoded frames of the same three cameras before augmentation (JPEG quality 90), the input for re-rendering
obb_yolo/ce877e8237be/	The YOLO-OBB detections on the head-left camera for every frame (up to 32 per frame, confidence ≥ 0.15); manifest.json records the detector

frame_cache_tars/ and frame_cache_v2_tars/ each have a restore.sh that unpacks their tar files, and a _cache_meta.json.

LeRobot features

feature	shape	content
observation.images.cam_0, cam_1	480 × 640 × 3	head camera, left and right eye
observation.images.cam_2, cam_3	480 × 640 × 3	left and right wrist camera
observation.images.cam_0_ir … cam_3_ir	480 × 640 × 3	infrared images of the two wrist stereo cameras (RealSense D405)
observation.state.robot_q_current	36	root pose (7) + 29 joint angles
observation.state.hand_state	2	Dex1-1 gripper positions
observation.state.ee_state	12	both hands' pose as recorded by BitRobot (xyz + euler xyz each)
action.robot_q_desired	36	commanded root pose + joint angles
action.hand_cmd	2	gripper commands
action.ee_action	12	commanded hand poses

meta/episodes/ also records, for every episode, where it comes from (source_episode_index, source_frame_start, source_frame_end, source_task_index) and its curation label (curation_verdict, curation_reviewer, curation_reviewed_at, …).

How the images were rendered

All variants are made from this dataset's own frames; no outside images are used.

• The overlay uses the YOLO-OBB weights at commit 8221d0a (recorded under the repository's former name Team-RAMEN/IROS2026_RAMEN_Hara_yoloobb_upperpolicy), drawing boxes with confidence ≥ 0.30 and line width 2, saved as JPEG with 4:4:4 chroma.
• The 9 augmented variants use colour jitter, blur, sharpening, noise and JPEG re-compression on all cameras, random erasing, and on the head camera a small rotation (±5°) and shift (±5%).
• The settings are in data/bitrobot_lerobot_subtask_datasets/configs/unified_frame_cache.yaml of the code repository.

Usage

As a LeRobot dataset (robot data and videos only):

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("Team-RAMEN/ramen-g1-ikea-assembly-ramen-ori-insert-table-leg")

To train RAMEN-Ori, download data/, meta/ and the frame_cache_v2_tars/ you need, unpack them with restore.sh, and point model/ramen_ori/configs/data/unified_insert.yaml (this skill only) or unified_all6.yaml (all six skills) at the result.

Notes

• observation.state.ee_state is the value recorded by BitRobot. It uses a different frame and tool offset from the forward kinematics in our code (10–18 cm apart); our newer models compute the hand pose from the joint angles instead.
• The labels that chose the segments come from one reviewer; see the curation dataset for its known issues.
• Older revisions of this repository may hold earlier versions of the caches (made from the same source frames).

License and attribution

This dataset is released under CC BY 4.0.

It is adapted from BitRobot/G1_WBT_Dex1_Building-Children-Table by BitRobot (revision 4b5961c), licensed under CC BY 4.0, with segment times from BitRobot/2026-humanoid-ikea-assembly-challenge. Changes: we cut the episodes into segments of one skill, kept the segments chosen by our curation labels, and added the image caches and YOLO-OBB detections described above.