SO101-StackBlock_Ours_100epi

100 successful demonstrations of stacking RED → GREEN → BLUE blocks on a blue dish, collected on a single LeRobot SO-101 arm. Recorded via the AutoDataCollector pipeline (Phase 1 buffer-aware seeding + Phase 2 MI-based useful-OOD acquisition).

Task

Stack red, green, and blue blocks on the blue dish from bottom to top.

Per episode:

1. Pick red block → place on blue dish
2. Re-detect → pick green block → stack on red
3. Re-detect → pick blue block → stack on green

Dataset stats

Episodes	100
Frames	86,402
FPS	10
Avg episode length	~864 frames (≈ 86 sec)
Robot	SO-101 (5-DoF + gripper)
Cameras	top (RealSense overhead) + left_wrist (wrist OpenCV)
Video codec	H.264 (yuv420p, 640×480 @ 10fps)

Features

Key	Shape	Notes
observation.state	(6,)	joint positions (5 arm + gripper)
action	(6,)	target joint positions
observation.images.top	(480, 640, 3)	overhead RGB, H.264
observation.images.left_wrist	(480, 640, 3)	wrist RGB, H.264
observation.ee_pos.robot_xyzrpy	(6,)	end-effector pose
observation.gripper_binary	(1,)	binary gripper state
skill.*	various	per-skill subgoal & verification metadata
subtask.*	various	subtask boundary labels

Collection method

• Phase 1 (80 episodes): buffer-aware subgoal seeding — each episode perturbs subgoals to diversify reached states across 20 random initial seeds × 5 slots.
• Phase 2 (20 episodes): MI-based useful-OOD acquisition replays Phase 1 subgoals but varies trajectory via curobo plan_batch + MI scoring.
• All 100 episodes verified via VLM judge (gemini-3-flash-preview) returning TRUE.

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("CoRL2026-CSI/SO101-StackBlock_Ours_100epi")
print(ds.num_episodes, ds.num_frames, ds.fps)

sample = ds[0]
top = sample["observation.images.top"]      # (3, 480, 640) torch.uint8
wrist = sample["observation.images.left_wrist"]
action = sample["action"]                    # (6,) joint targets

Source

Collected with AutoDataCollector (CoRL 2026 submission).