rocochallenge2025/rocochallenge2026_Industrial_Assembly

RoCo Challenge 2026 — Industrial Assembly (vega_1u)

A LeRobot v3.0 dataset of bimanual pick-and-place task-board assembly demonstrations, collected in NVIDIA Isaac Sim 5.1.0 on the IROS 2026 RoCo Challenge taskBoardAssembly environment with the vega_1u dual-arm robot.

Every episode is one complete 9/9 successful assembly: all nine parts are picked and placed onto the task board in a fixed order. Part positions and the assembly order are held constant across episodes; the robot's motion trajectories are randomized, so the demonstrations are diverse while always succeeding.

At a glance

Format	LeRobot v3.0
Robot	vega_1u (dual-arm, 7-DOF per arm)
Control / sample rate	10 fps
Episodes	200 (each a full 9-part assembly)
Frames	121,454 (~607 frames ≈ 60 s per episode; range 593–715)
Cameras	3 × RGB 240×320×3, H.264 video
State dim	44
Action dim	14
Single task	"assemble parts onto the task board"
Size on disk	~1.1 GB (images stored as H.264 video; ~84 GB if raw)

Features

Key	Type	Shape	Notes
observation.images.head	video	(240, 320, 3)	head/overhead camera, RGB uint8
observation.images.left_hand	video	(240, 320, 3)	left wrist camera
observation.images.right_hand	video	(240, 320, 3)	right wrist camera
observation.state	float32	(44,)	see layout below
action	float32	(14,)	see layout below

Standard LeRobot bookkeeping columns (timestamp, frame_index, episode_index, index, task_index) are also present.

observation.state (44) layout

Concatenated in this exact order:

Slice	Dims	Field	Units / convention
[0:7]	7	left arm end-effector pose	x, y, z (m, world frame) + quaternion qw, qx, qy, qz
[7:14]	7	right arm end-effector pose	same convention
[14:21]	7	left arm joint positions	radians (7 joints)
[21:28]	7	right arm joint positions	radians
[28:35]	7	left arm joint velocities	rad/s
[35:42]	7	right arm joint velocities	rad/s
[42:43]	1	left gripper position	open ratio [0,1] (1 = fully open)
[43:44]	1	right gripper position	open ratio [0,1]

action (14) layout

Target end-effector command, left arm then right arm (7 each):

Slice	Dims	Field	Units / convention
[0:3]	3	left EE target position	x, y, z (m, world frame)
[3:6]	3	left EE target orientation	intrinsic XYZ Euler (rad), see note
[6:7]	1	left gripper command	open ratio [0,1]
[7:10]	3	right EE target position	x, y, z (m)
[10:13]	3	right EE target orientation	intrinsic XYZ Euler (rad)
[13:14]	1	right gripper command	open ratio [0,1]

Euler note: action orientation columns are unwrapped over time (np.unwrap) to remove ±π discontinuities, so values can exceed ±π (observed range up to ~7 rad). This is intentional — it gives a continuous regression target. Wrap back to (-π, π] if your controller needs it.

Single-arm baseline note: these demonstrations are generated by a baseline policy that solves the task primarily with the left arm; the right arm holds a near-constant home pose throughout. Right-arm state and action columns therefore vary very little. The robot itself is fully bimanual.

Coordinate & unit conventions

• Positions in meters, world frame.
• EE pose orientation in observation.state is a quaternion in (w, x, y, z) order; action orientation is XYZ-Euler radians. They describe the same EE but in different parameterizations — convert as needed.
• Joint angles/velocities in radians / rad/s; vega_1u arms are 7-DOF.
• Gripper values are normalized open ratios in [0,1] (physical joint travel is [0, 0.6649704] rad; ratio = value / 0.6649704).

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("rocochallenge2025/rocochallenge2026_Industrial_Assembly")
print(ds.num_episodes, ds.num_frames)   # 200  121454

sample = ds[0]
sample["observation.images.head"].shape   # torch.Size([3, 240, 320])
sample["observation.state"].shape         # torch.Size([44])
sample["action"].shape                    # torch.Size([14])

Normalization statistics live in meta/stats.json (use LeRobot's make_pre_post_processors or your own normalizer). Images are decoded from H.264 video on the fly, so you get full-quality frames at the original 240×320 resolution.

How it was collected

• Environment: IROS 2026 RoCo Challenge taskBoardAssembly, Isaac Sim 5.1.0, deterministic PhysX (enhanced determinism + single-thread solver).
• Policy: a randomized baseline pick-and-place policy. Assembly order and part positions are fixed; per-episode motion trajectories are perturbed for diversity (mean cross-episode action std ≈ 0.10 on left-arm dims).
• Filtering: only fully successful 9/9 rollouts are kept; partial assemblies are discarded.
• Sampling: one row per simulation control step at 10 Hz.

Integrity

This release was validated before upload (v3.0 compliant; 200 episodes / 121,454 frames consistent across info.json, parquet, and videos; no NaN/Inf; finite stats with no zero-std dims; all 200 episodes decode and frame-align; all files byte-for-byte size-matched on the Hub). Low-dim value ranges: action ∈ [-3.22, 7.00], state ∈ [-14.58, 36.64].

License

Released under the MIT License.

Citation

If you use this dataset, please cite the IROS 2026 RoCo Challenge.