ShuaKang/my_roboTwin2.0_training

RoboTwin 2.0 — coverage-controlled training data

Simulated bimanual manipulation demonstrations collected in RoboTwin 2.0, exported to LeRobot v2.1. The point of this collection is a co-training / domain-coverage ablation: five pretraining sets that differ only in how much visual and spatial variation the scene generator was allowed to sample, plus two small downstream target tasks to fine-tune and evaluate on.

RoboTwin 2.0 is a scalable simulation benchmark for bimanual manipulation (dual-arm AgileX / ALOHA-style robot). Its domain-randomisation pipeline procedurally swaps scene backgrounds, drops in clutter objects, perturbs object placement, and generates many natural-language phrasings per task, which is exactly what is being dialled up and down here.

Collection code: https://github.com/Shua-Kang/my_roboTwin2.0

---

The three randomisation axes

Every pretraining variant uses the same task (place_a2b_right), the same robot, the same controller, and the same episode budget. They differ only along three knobs:

variant	background textures	distractor classes	work area (object x)
pre_full	100	10	full range
pre_none	1 (none)	0	x fixed
pre_bg	100	0	x fixed
pre_dt	1 (none)	10	x fixed
pre_wa	1 (none)	0	full range

pre_full is all three knobs on, pre_none is all three off, and the three single-axis variants isolate one knob each (BG = background, DT = distractors, WA = work area).

Counts measured on the released episodes

These are not the config values, they are what the shipped data actually contains:

measured quantity	pre_full	pre_none	pre_bg	pre_dt	pre_wa
distinct background textures	100	0	99	0	0
distinct distractor classes	8	0	0	8	0
mean distractors per episode	1.50	0	0	1.49	0
object x-span (m)	0.206	0.027	0.033	0.036	0.193

The realised distractor-class count is 8 rather than the configured 10 (two classes were never sampled in 250 episodes), and pre_bg realised 99 of 100 textures. The "x fixed" variants still show a 2.7–3.6 cm span because placement is jittered even when the work-area knob is off; the "full range" variants are ~6x wider at ~0.20 m.

---

Camera setup (v2, regenerated 2026-09-08)

All datasets were re-rendered on 2026-09-08 with a new head-camera placement; the trajectories, seeds, scene layouts and actions are bit-identical to v1, only the observation.images.head stream (and its intrinsics/extrinsics) changed.

	v1 (old)	v2 (this release)
head camera position (m)	(-0.032, -0.45, 1.35)	(0.10, -0.65, 1.50)
pitch (down from horizontal)	53.1°	52°
vertical FOV	37°	43°
resolution	320x240	320x240
robot body / mast / left arm	visible	hidden (render only)

Why: with the v1 camera the right arm at its home pose is out of frame, so the opening frames of every episode show no robot at all. The v2 camera sits further back over the robot base so the whole right arm is visible from frame 0. Pulling the camera back would put the robot mast and the idle left arm in view, so every link except the right arm is made invisible (RenderBodyComponent.visibility = 0; physics and collision are untouched — the left arm is still there, it just is not drawn). The right-wrist camera is unchanged.

In RoboTwin task-config terms (see the collection repo): camera.head_camera_type: D435_fov43, camera.head_camera_pose: {position: [0.10, -0.65, 1.50], forward: [0, 0.6157, -0.7880], left: [-1, 0, 0]}, hide_robot_links: all_but_right. The evaluation configs use the same settings, so policies trained on this data see the same rendering at test time.

---

Datasets in this repo

Each directory is one self-contained LeRobot v2.1 dataset under <name>/lerobot_delta_hr/ (hr = head + right-wrist cameras, delta = end-effector delta actions).

Pretraining — task place_a2b_right

path	episodes	frames
pre_full_place_a2b_right/lerobot_delta_hr	250	36,991
pre_none_place_a2b_right/lerobot_delta_hr	250	36,929
pre_bg_place_a2b_right/lerobot_delta_hr	250	37,099
pre_dt_place_a2b_right/lerobot_delta_hr	250	36,968
pre_wa_place_a2b_right/lerobot_delta_hr	250	37,191

Downstream targets

path	episodes	frames
narrow_place_container_plate/lerobot_delta_hr	50	7,823
narrow_move_pillbottle_pad/lerobot_delta_hr	50	7,324

Both targets are deliberately small (50 demos), narrow-coverage sets — the thing the pretraining variants are meant to help with.

---

Format

• LeRobot v2.1 (meta/info.json → codebase_version: v2.1), robot_type: aloha-agilex, 30 fps.
• data/chunk-000/episode_XXXXXX.parquet — one parquet per episode.
• videos/chunk-000/<video_key>/episode_XXXXXX.mp4 — two streams per episode, observation.images.head and observation.images.right_wrist, both 320x240 (shape: [240, 320, 3]), h264 / yuv420p, no audio.
• meta/ — info.json, stats.json, episodes.jsonl, episodes_stats.jsonl, tasks.jsonl, and a GR00T-style modality.json giving the sub-slices of the state and action vectors.

action — 14-D, end-effector delta

Per arm, [dx, dy, dz, rotvec(3), gripper], left arm first:

dims	meaning
0:3	left EE translation delta
3:6	left EE rotation delta (rotation vector / axis-angle)
6	left gripper (absolute, 0 closed … 1 open)
7:10	right EE translation delta
10:13	right EE rotation delta (rotation vector)
13	right gripper (absolute)

observation.state — 20-D, absolute

Per arm, [pos(3), rot6d(6), gripper(1)], left arm first: dims 0:10 left, 10:20 right. The rotation is the 6D continuous representation (first two columns of the rotation matrix), not Euler angles or a quaternion.

Language

annotation.human.task_description and annotation.human.task_name are integer indices into meta/tasks.jsonl, stored per frame. Index 0 is the coarse task name (e.g. place a2b right); the description index rotates frame by frame through all of that episode's phrasings and wraps around, so a 149-frame episode cycles through ~99 different sentences describing the same goal. Do not assume one sentence per episode — sample the frame's own index, or pool the episode's set, depending on what you want.

Example phrasings for one place_a2b_right episode:

Use the right arm to place the rubikscube featuring white, green, orange tiles right of the curved black stapler. Set the puzzle rubikscube with 3x3 grid per side carefully on the black stapler with rectangular base's right side.

---

Important gotchas

Only the right arm moves

Every episode in this collection is pinned to the right arm. The left-arm action dimensions are therefore constant: action[0:6] is exactly zero in every frame and action[6] (left gripper) is pinned at 1.0. observation.state[0:10] is likewise the fixed left-arm home pose (per-dimension std ≤ 2.5e-4). If you train on all 14 action dims, six of them carry no signal and one is a constant — normalise accordingly, or slice out action[7:14].

Integrate deltas on the previously commanded pose

The actions are deltas, and the reference frame matters. Accumulate each delta on the pose you last commanded, not on a freshly observed one. Re-basing on the observation folds the controller's steady-state tracking error into every step; it compounds and drifts roughly 58 mm per episode, which is enough to miss a grasp entirely. Keep an internal commanded-pose target, add the predicted delta to it, and send that.

# correct
cmd = cmd + delta          # cmd initialised once from the first observation
# wrong — bleeds tracking error every step
cmd = observe_ee() + delta

---

Loading

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset(
    "ShuaKang/my_roboTwin2.0_training",
    root="pre_full_place_a2b_right/lerobot_delta_hr",   # or download the subfolder locally
)

or just pull one subfolder:

hf download ShuaKang/my_roboTwin2.0_training --repo-type dataset \
    --include "pre_full_place_a2b_right/lerobot_delta_hr/*" --local-dir ./robotwin_cov

Because this repo holds several independent LeRobot datasets side by side, each subfolder must be treated as its own dataset root — there is no single top-level meta/.

---

Citation

RoboTwin 2.0 — please cite the upstream benchmark if you use this data. Collection and export code for these particular sets: https://github.com/Shua-Kang/my_roboTwin2.0