RoboticArmKnotTying-G3

This dataset contains one complete teleoperated demonstration in which two AgileX Piper arms tie a knot in a red rope. Group G3 recorded it on 2026-10-13 for the course 机器人具身智能系统开发实践基础, Part 2 (data collection).

• Task: Tie a knot in the red rope using both arms
• Format: LeRobot v3.0, written with lerobot 0.6.0
• Size: 1 episode, 2608 frames at 30 fps (86.9 s)
• Robot: agilex_piper_dual. Two leader/follower pairs. The left arm is on can0 and the right arm is on can1. Each leader and its follower share one CAN bus at 1 Mbps.
• Cameras: three Intel RealSense D435 cameras, each recording RGB and depth at 640×480 and 30 fps. Depth is aligned to the colour image.
  • cam_high: fixed top-down view (serial 261322071786)
  • cam_left_wrist: left wrist (serial 261322074905)
  • cam_right_wrist: right wrist (serial 346522076815)

Features

key	shape	content
observation.state	14	Follower joint feedback. Per arm: 6 joint angles (rad) and the gripper opening (m). Order is [left…, right…].
action	14	Leader joint commands (CAN frames 0x155–0x159), in the same layout as observation.state.
observation.eef_pose	14	Follower end pose. Per arm: x y z (m, in the arm's base frame) and a quaternion qx qy qz qw. The source is GetArmEndPoseMsgs.
action.eef_pose	14	Leader end pose, computed by forward kinematics of the leader joints (piper_sdk, dh_is_offset=1).
observation.velocity	14	Follower joint velocities (rad/s), from the high-speed feedback.
observation.effort	14	Follower joint torques and gripper effort.
observation.images.<cam>	480×640×3	RGB video (AV1).
observation.images.<cam>_depth	480×640×1	Depth video, lossless. It is encoded as HEVC 12-bit with linear quantization over 0–4.095 m, so each level is exactly 1 mm. The decoded values equal the raw uint16 millimetres, and 0 means invalid.
task	str	Natural-language task description.

meta/piper_extra_info.json holds:

• the intrinsics (K, D, depth scale, serial) of every camera;
• the field layouts;
• quality statistics;
• the hand-eye calibration of the left wrist camera (eye-in-hand, T_link6_cam):
  • translation (−75.5, −0.2, 37.9) mm;
  • board-in-base residual 3.9 mm / 1.1°, using 15 valid poses.

Data quality

check	result
sampling interval	33.3 ms mean (30.5–36.1 ms)
repeated frames per stream	≤ 1.2 %
cross-stream timestamp skew	57 ms mean, 104 ms max
leader→follower tracking	0.016 rad RMS (left), 0.020 rad RMS (right), lag ≈ 3 frames
gripper range	0.7–55.1 mm (left), −3.1–56.6 mm (right)
static head / tail	0.13 s / 0.5 s, no trimming needed
valid depth pixels	91 % (high), 70 % (left wrist), 41 % (right wrist: too close to the rope for the D435 minimum range)

The episode starts with the rope lying loose on the table and ends with a knot tied in the middle of the rope. You can see the knot in the last cam_high frame.

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("KeweiXu/RoboticArmKnotTying-G3")
frame = ds[0]
frame["task"], frame["observation.state"], frame["observation.images.cam_high_depth"]  # depth in mm

We collected the data with the configuration-driven ROS 2 pipeline from our course report, adapted from cobot_ros2_datacollect_PKU. The raw HDF5 episode (12 GB) has the same fields, plus per-stream timestamps.