yuchensong/glide_data

GLIDE:
Learning Beyond What Humans Can Demonstrate

Yuchen Song , Aditya Mittal , Unnat Jain

TL;DR: GLIDE uses closed-loop, self-refined guardrails to make infeasible robot manipulation tasks demonstrable, learnable, and the trained policies more reliable to deploy.

This repository hosts the real-robot demonstration datasets used to train the GLIDE π₀.₅ policies. The robot runtime used to collect the data and evaluate policies is in glide_code.

Tasks

• 🍅 Tomato plate transfer (lift_plate_*): use both grippers to grasp a plate of tomatoes and carry it from the tabletop to an elevated surface without tilting the plate or spilling.
• 🖍️ Marker handover & stand (stand_marker_*): grasp a marker with one gripper, transfer it to the other gripper, place it upright on the table, and withdraw without knocking it over.
• 🍷 Wine serving (pour_wine_*): use a left parallel-jaw gripper to hold a bottle and a right 15-DoF CRAFT hand to hold a glass while pouring wine from the bottle into the glass.

Subsets

Each task has two subsets, each a self-contained LeRobot dataset:

• *_mixed: every recorded demonstration, successful and failed (mixed quality).
• *_success: only the successful episodes from the matching *_mixed subset.

Subset	Robot	Episodes	Frames	FPS	Duration	Size
lift_plate_mixed	yam_dual_arm	30	13,170	30	7.3 min	144 MB
lift_plate_success	yam_dual_arm	19	8,508	30	4.7 min	94 MB
stand_marker_mixed	yam_dual_arm	30	12,769	30	7.1 min	128 MB
stand_marker_success	yam_dual_arm	22	9,216	30	5.1 min	93 MB
pour_wine_mixed	yam_dual_arm_craft_hand	60	162,158	45	60.1 min	2.3 GB
pour_wine_success	yam_dual_arm_craft_hand	30	83,240	45	30.8 min	1.2 GB

Language instructions (one per task):

Task	Instruction
🍅 Tomato plate transfer	"Use both arms to pick up the plate and put it onto the box"
🖍️ Marker handover & stand	"Pick up the marker with one gripper, transfer it to the other gripper, and lift it upright on the table."
🍷 Wine serving	"Use left gripper to pick up the wine bottle, right hand to pick up the wine cup, and pour wine from bottle to the cup."

Data Format

All subsets use the LeRobot dataset format, codebase_version: v2.1:

<subset>/
├── meta/
│   ├── info.json              # schema, fps, feature names
│   ├── tasks.jsonl            # language instruction
│   ├── episodes.jsonl         # per-episode length and task
│   └── episodes_stats.jsonl   # per-episode feature statistics
├── data/chunk-000/episode_XXXXXX.parquet
└── videos/chunk-000/observation.images.{head,left_wrist,right_wrist}/episode_XXXXXX.mp4

Every subset records three RGB cameras (head, left_wrist, right_wrist) at 640×480 as H.264 MP4, along with the robot state, the executed action, and the raw teleoperation signals.

Two-gripper tasks (lift_plate_*, stand_marker_*)

Two YAM arms, each with a parallel-jaw gripper, teleoperated with Quest controllers.

Feature	Shape	Description
observation.state	14	left_joint_0..5, left_gripper, right_joint_0..5, right_gripper
action	14	Commanded joints and grippers, same order as the state
teleoperation.matrices.{head,left_controller,right_controller}	4×4	Quest head and controller poses
teleoperation.buttons	8	Trigger, squeeze, A/B buttons for each controller

CRAFT-hand task (pour_wine_*)

A left YAM arm with a parallel-jaw gripper and a right YAM arm with a 15-DoF CRAFT hand, teleoperated with Quest hand tracking.

Feature	Shape	Description
observation.state	28	left_joint_0..5, left_gripper, right_joint_0..5, craft_motor_0..14_raw
action	28	Commanded joints, gripper, and CRAFT motors, same order as the state
teleoperation.matrices.{head,left_hand,right_hand}	4×4	Quest head and wrist poses
teleoperation.landmark_matrices.{left_hand,right_hand}	25×4×4	Quest hand-landmark poses
teleoperation.hand_states	12	Pinch, squeeze, and tap flags and values per hand
teleoperation.tracking	9	Tracking validity, event counts, and packet ages for head and hands

All subsets also include the standard LeRobot index columns: timestamp, frame_index, episode_index, index, and task_index.

Usage

Download a single subset:

hf download yuchensong/glide_data \
  --repo-type dataset \
  --include "stand_marker_success/*" \
  --local-dir ./glide_data

Load it with a LeRobot release that reads the v2.1 format, such as the version pinned by OpenPI:

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset(
    repo_id="yuchensong/glide_data",
    root="./glide_data/stand_marker_success",
)
frame = dataset[0]
print(frame["observation.state"].shape, frame["action"].shape)

Or read the tabular data directly:

import pandas as pd

df = pd.read_parquet(
    "./glide_data/stand_marker_success/data/chunk-000/episode_000000.parquet"
)

To train π₀.₅ policies, follow OpenPI's instructions for computing normalization statistics and fine-tuning on a local LeRobot dataset.

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Citation

@article{song2026glide,
  title={Learning Beyond What Humans Can Demonstrate},
  author={Song, Yuchen and Mittal, Aditya and Jain, Unnat},
  journal={arXiv preprint arXiv:2609.24996},
  year={2026}
}