真机 SO-101 主从臂遥操采集的抓橙子放盘子数据集,30 集人类演示,LeRobot v3.0 格式。 Real-world SO-101 leader-follower teleoperation dataset: pick up an orange and place it on the plate, 30 human demonstrations in LeRobot v3.0 format.

概览 / Overview

项	值
机器人 / Robot	SO-101 follower(6 DoF + 夹爪,leader 臂遥操)
任务 / Task	`Pick up the orange and place it on the plate`
集数 / Episodes	30
总帧数 / Frames	25,091(30 fps,合计 ~14 min)
单集长度 / Episode length	442–2,195 帧(median 753 ≈ 25 s)
相机 / Cameras	`front` + `wrist`,各 640×480 @30fps,AV1(yuv420p)
状态与动作 / State & action	6 维关节位置(shoulder_pan / shoulder_lift / elbow_flex / wrist_flex / wrist_roll / gripper)
格式 / Format	LeRobot `codebase_version: v3.0`

采集方式 / Collection

主从臂遥操(leader 5V / follower 12V),人类演示自然速度完成抓取-放置全流程。 Leader-follower teleoperation; each episode is a complete human-paced pick-and-place.
录制脚手架(后台录制 / 哨兵停止 / 废集丢弃)见下方 GitHub 仓库 LeIsaac/scripts/so101/。 Recording scaffolding (background record, sentinel stop, drop-bad-episode) lives in the GitHub repos below.

用法 / Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("wsagi/leisaac-real-pick-orange")
print(ds.meta.total_episodes, ds.meta.total_frames)  # 30, 25091

⚠️ 视频为 AV1 编码:decord 读不了,torchvision_av/pyav 后端可读(pyav 建议 thread_count=1 防解码线程泄漏)。 Videos are AV1-encoded: use the pyav / torchvision_av backend (decord cannot decode AV1).

在此数据集上训练的结果 / Trained on this dataset

同一 30 集数据训练三种 VLA 头,真机闭环对比: Three policy heads trained on the same 30 episodes, closed-loop on the real arm:

每轮 3 只橙子,记放入盘中数 / 3 oranges per round, count placed:

策略 / Policy	真机表现 / Real-arm result
FlowDP sim-warm-start(wsagi/FlowHeads-DiffusionPolicy-Real-PickOrange)	1.7 只/轮 = 56.7%(10 轮:2,2,3,1,1,2,0,3,2,1;超自身仿真成绩 45.0%)— 真机全场 best / real-arm best, beats its own sim score
ACT sim-warm-start(wsagi/ACT-Real-PickOrange)	1.5 只/轮(10 轮:2,1,1,3,2,1,0,1,2,2)/ ~2× from-scratch
ACT 从零 / from scratch	0.8 只/轮(5 轮:2,0,1,1,0;放入 1–2 只概率约 50% / ~50% chance of placing 1–2)
FlowDP 从零 / from scratch	能抓取,放置不稳 / grasps, unstable placement
GR00T-N1.7(冻结 VLM 微调 head)	能抓取,未完成放置 / grasps, no placement