2shoe-v2: "Place 2 Shoes" demonstrations

Teleoperated demonstrations for the multi-instance shoe task (RQ5, "Place 2 Shoes") in:

T. Lips, M. Moletta, M. C. Welle, D. Kragic, F. wyffels. On the Generalization Capabilities, Design Choices and Limitations of Keypoint Imitation Learning. arXiv:2605.26649, 2026. paper · project page

Recorded by Thomas Lips (UGent AIRO) on 2026-02-03. Uploaded as-is from the original recording, so it is not in LeRobot format. It uses the incar recording format (see below).

Demos	51 (demo_0 … demo_50), none are teleop takeovers
Demo length	27.6 – 48.7 s (mean 32.5 s)
Size	~69 GB (mostly depth maps)
Arm	single (right) arm, 7-D pose + gripper
Cameras	Orbbec static scene camera (RGB + depth), OAK wrist camera

Layout

dataset_config.json              # feature list (types/shapes) used by the incar recorder
demo_N/
  meta.json                      # t_start, t_end, is_takeover, per-feature statistics (min/max/mean/std/count)
  orbbec_scene/                  data.mp4 + timestamps.h5    RGB, 640x480, ~30 fps
  orbbec_scene_depth_image/      data.mp4 + timestamps.h5    colourised depth (visualisation only)
  orbbec_scene_depth_map/        data.h5 ["data"] (T,480,640) float32, metres; ["timestamps"]
  oak_right_wrist/               data.mp4 + timestamps.h5    wrist RGB, 640x480, ~30 fps
  robot_state_right/             data.h5 ["data"] (T,7) float64, ~50 Hz; ["timestamps"]
  teleop_action_right/           data.h5 ["data"] (T,7) float64, ~10 Hz; ["timestamps"]
  buttons_right/                 data.h5 ["data"] JSON strings of Quest controller buttons; ["timestamps"]
*.json                           # Thomas's training configs for this dataset (see below)
keypoint_configs/                # reference keypoints used by the KIL preprocessing (see below)
camera_calibration/              # Orbbec scene camera intrinsics + extrinsics

Streams and synchronisation

• Every stream has its own timestamps array (int64, microseconds), recorded at its own rate. Streams have to be aligned by timestamp. Thomas's pipeline resamples observations to the action timestamps (sample_observations_to_action_timestamps, 10 Hz, dt = 0.1).
• The timestamps look like microseconds since midnight, and some streams start just before midnight (e.g. 86399998513) and wrap around to 0. Unwrap them modulo 86 400 000 000 before aligning.
• robot_state_right is [x, y, z, roll, pitch, yaw, gripper]: TCP position in metres in the robot base frame, Euler angles in radians, gripper opening. teleop_action_right uses the same 7-D absolute target pose + gripper convention. For training, Thomas converts the action orientation to the 6-D rotation representation, giving a 10-D action (pose_euler_to_6d_orientation).
• buttons_right contains the Quest controller state. The filter_by_buttons step uses it to drop frames where the operator was not engaged, and the first second of actions is also cut (cut_first_n_action_steps).

Reading a demo

import h5py, cv2

d = "2shoe-v2/demo_0"
with h5py.File(f"{d}/robot_state_right/data.h5") as f:
    state, state_t = f["data"][:], f["timestamps"][:]
with h5py.File(f"{d}/teleop_action_right/data.h5") as f:
    action, action_t = f["data"][:], f["timestamps"][:]
with h5py.File(f"{d}/orbbec_scene/timestamps.h5") as f:
    rgb_t = f["timestamps"][:]
cap = cv2.VideoCapture(f"{d}/orbbec_scene/data.mp4")   # frame i <-> rgb_t[i]
with h5py.File(f"{d}/orbbec_scene_depth_map/data.h5") as f:
    depth0 = f["data"][0]                                 # (480, 640) metres, aligned to the RGB frame

Download only part of the data (the depth maps are most of the size):

huggingface-cli download ramen-noodels/2shoe-v2 --repo-type dataset --local-dir 2shoe-v2 \
  --include "demo_*/orbbec_scene/*" "demo_*/oak_right_wrist/*" "demo_*/robot_state_right/*" \
            "demo_*/teleop_action_right/*" "demo_*/buttons_right/*" "demo_*/meta.json" "*.json" "keypoint_configs/**" "camera_calibration/*"

Keypoints and calibration

• keypoint_configs/<name>/config.json lists the objects and their reference keypoints (pixel coordinates in a reference image in img/). KIL finds the corresponding points in every frame with a visual foundation model and lifts them to 3-D using the depth map and the camera calibration.
  • 1shoe-tray-v2-kp: tray (3 kp) + shoe (6 kp). This is the config the 2-shoe policies use, with detect_instances: true (SAM3, max_num_instances: 2), so the shoe keypoints are found on both shoe instances.
  • 2shoe-tray-v1-kp: tray (3) + shoe (6) + shoe2 (6), an alternative config with two explicitly separate shoe objects.
• camera_calibration/: camera_intrinsics.json is the one used in the training configs (fx = fy = 353.84, cx = 265.72, cy = 236.58, 640x480). camera_extrinsics.json gives the camera pose in the robot base frame (translation in m, XYZ Euler angles in degrees). camera_intrinsics_no_crop.json is the uncropped variant.

Training configs (*.json at the root)

These are Thomas's incar training configs for the methods compared in the paper. File paths in them point to his machine (/home/tlips/incar_ws/...) and must be changed.

• SKIL-M-*, SKIL-L-*, KAT-*: keypoint policies (KIL variants). The name gives the keypoint model (e.g. sam3-DIFTe2 = SAM3 instances + DIFT features, radioB = RADIO-B), and suffixes give the augmentation used. The main 2-shoe config is SKIL-M-sam3-DIFTe2-instance-aug.json (keypoint diffusion policy, horizon 16, 12 action steps, 2 obs steps, spatial and instance-reordering augmentation).
• dp-rgb-train_config*.json: RGB diffusion policy baseline.
• s2_diffusion.json: S2-diffusion baseline.

Code

The recorder, preprocessing (keypoint extraction) and training code is Thomas's incar codebase (INCAR-Robotics/core), which is not public. Ask Thomas Lips for access. Without it you can still read the raw data with h5py + any video decoder, as shown above.

Contact

Thomas Lips (author / data) · Raman Talwar (uploader), UGent AIRO.