CoRL2026-CSI/SO101-StackBlock_Ours_100epi

SO101-StackBlock_Ours_100epi数据集包含100个成功的演示，展示了在蓝色盘子上从底部到顶部堆叠红色、绿色和蓝色积木块的任务。这些数据使用单个LeRobot SO-101机械臂收集，通过AutoDataCollector管道记录（第一阶段缓冲区感知播种+第二阶段基于互信息的有用OOD获取）。每个演示包括：拾取红色积木块→放置在蓝色盘子上；重新检测→拾取绿色积木块→堆叠在红色积木块上；重新检测→拾取蓝色积木块→堆叠在绿色积木块上。数据集统计：86,402帧，每秒10帧，平均每集约864帧（约86秒）。机器人使用SO-101（5自由度+夹爪）。摄像头包括顶部（RealSense overhead）和左腕部（wrist OpenCV）。视频编解码器为H.264。特征包括观察状态（关节位置）、动作（目标关节位置）、图像（顶部和腕部RGB）、末端执行器位姿、夹爪状态、技能元数据和子任务边界标签。收集方法：第一阶段（80集）使用缓冲区感知子目标播种；第二阶段（20集）使用基于互信息的有用OOD获取。所有演示通过VLM法官（gemini-3-flash-preview）验证返回TRUE。SO101-StackBlock_Ours_100epi is a dataset containing 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). Each episode involves: picking red block → placing on blue dish; re-detecting → picking green block → stacking on red; re-detecting → picking blue block → stacking on green. Dataset stats: 86,402 frames, 10 FPS, average episode length ~864 frames (≈86 seconds). Robot: SO-101 (5-DoF + gripper). Cameras: top (RealSense overhead) and left_wrist (wrist OpenCV). Video codec: H.264. Features include observation.state (joint positions), action (target joint positions), observation.images.top and observation.images.left_wrist (RGB images), observation.ee_pos.robot_xyzrpy (end-effector pose), observation.gripper_binary (gripper state), skill.* metadata, and subtask.* labels. Collection method: Phase 1 (80 episodes) uses buffer-aware subgoal seeding; Phase 2 (20 episodes) uses MI-based useful-OOD acquisition. All episodes verified via VLM judge (gemini-3-flash-preview) returning TRUE.