Dataset Card for MotIF-1K

MotIF-1K is a robotics motion dataset containing 1,022 demonstrations across 13 task categories, used to benchmark and fine-tune vision-language models (VLMs) for motion-based success detection. Each demonstration includes a video of the motion, multiple pre-rendered trajectory visualizations, task instructions, and motion descriptions.

The FiftyOne dataset is a grouped dataset where each group represents one trajectory and each group slice represents a different visual representation of that trajectory — mirroring the exact input formats used in the paper.

This is a FiftyOne dataset with 1023 samples.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("Voxel51/motif-1k")

# Launch the App
session = fo.launch_app(dataset)

Dataset Details

Dataset Sources

• HuggingFace repository: https://huggingface.co/datasets/myconnects/motif
• Paper: https://arxiv.org/abs/2409.10683
• Code / collection scripts: https://github.com/Minyoung1005/motif
• Paper: MotIF: Motion Instruction Fine-tuning (Hwang et al., 2024)
• Project page: https://motif-1k.github.io
• License: MIT

---

FiftyOne Dataset Structure

Grouped Dataset Overview

Dataset name:   motif-1k
Media type:     group
Default slice:  video_trajviz
Groups:         1,022  (653 human_motion + 369 stretch_motion)

Group Slices

Every trajectory group contains up to 13 slices. Each slice is a separate fo.Sample with its own media file and labels. Not all slices are present for every group — see the Incomplete samples note below.

Slice name	Media type	Description	Always present?
video_trajviz	video	Raw video with the trajectory overlaid — the default slice and the paper's primary representation	No (absent for 182 incomplete stretch samples)
video_raw	video	Clean video without any trajectory overlay; carries the interactive per-frame trajectory Polyline	Yes
last_frame_trajviz	image	Final video frame with trajectory overlay — the exact image input used by the paper's VLM	No
last_frame_raw	image	Final video frame, no overlay	No
opticalflow	image	Full optical-flow visualization of all keypoints	No
storyboard_key2	image	2-keyframe storyboard, clean	No
storyboard_key2_trajviz	image	2-keyframe storyboard with trajectory overlay	No
storyboard_key4	image	4-keyframe storyboard, clean	No
storyboard_key4_trajviz	image	4-keyframe storyboard with trajectory overlay	No
storyboard_key9	image	9-keyframe storyboard, clean	No
storyboard_key9_trajviz	image	9-keyframe storyboard with trajectory overlay	No
storyboard_key16	image	16-keyframe storyboard, clean	No
storyboard_key16_trajviz	image	16-keyframe storyboard with trajectory overlay	No

Sample-Level Fields

All fields below are present on every slice of every group.

Field	Type	Description
group	Group	FiftyOne group handle; group.id is the trajectory identifier, group.name is the slice name
config	str	Source config: "human_motion" or "stretch_motion"
traj_idx	int	Trajectory index within its config (0-based)
task_instruction	str	High-level task goal, e.g. "shake the boba"
motion_description	str	Fine-grained motion specification, e.g. "move to the right and to the left, repeating this sequence 3 times"
num_steps	int	Number of steps as stored in the source (may differ from trajectory_length; see notes)
trajectory_length	int	Actual number of trajectory points (len(trajectory)) — the reliable count
has_source_artifacts	bool	Whether this sample's group has all pre-rendered visualizations. False for 182 incomplete stretch_motion groups
tags	list[str]	Always includes the config name; incomplete groups are also tagged "incomplete"

Label Fields

video_raw slice — frames.trajectory (per-frame Polyline)

The video_raw slice carries a frame-level progressive trajectory annotation. At frame N, the Polyline contains the first N trajectory points, so the path draws itself out as the video plays.

• Frame 1: a zero-length degenerate segment marking the trajectory start position (renders as a dot)
• Frame N: the full trajectory path accumulated to that point

Each Polyline carries these label attributes:

Attribute	Type	Description
coord_space	str	Coordinate convention used: video_pixels, video_pixels_offset, or realsense_native
has_source_artifacts	bool	Whether the source provided a last_frame_trajviz for offset detection
correction_method	str	How the trajectory was corrected: identity, detected, resolution_median_fallback, default_fallback, or realsense_heuristic
offset_x	float	Pixel offset applied in x (0 for identity and realsense_heuristic)
offset_y	float	Pixel offset applied in y (0 for identity and realsense_heuristic)

All Polyline coordinates are normalized to [0, 1] × [0, 1] relative to the video frame.

---

Dataset Composition

Config	Agent	Trajectories	Has all slices?
human_motion	Human (6 different people)	653	Yes — all 13 slices
stretch_motion (with artifacts)	Hello Robot Stretch 2	188	Yes — all 13 slices
stretch_motion (incomplete)	Hello Robot Stretch 2	182	video_raw only; tagged "incomplete"
Total		1,022

Task Categories

13 categories spanning non-interactive, object-interactive, and user-interactive motions:

Category	Tasks
Non-interactive	Outdoor Navigation, Indoor Navigation, Draw Path
Object-interactive	Shake, Pick and Place, Stir, Wipe, Open/Close Cabinet, Spread Condiment
User-interactive	Handover, Brush Hair, Tidy Hair, Style Hair

---

Trajectory Coordinate System

The trajectory field in the source data stores 2D pixel coordinates [x, y] per timestep. The coordinate space differs by config — this is a known source-side inconsistency, not a parsing bug:

coord_space value	Applies to	Correction applied
video_pixels	All human_motion (653)	Identity — MediaPipe hand detection runs on the cropped video frame, so coordinates match the stored video dimensions directly
video_pixels_offset	stretch_motion with artifacts (188)	Per-sample pixel translation detected from the red endpoint marker in last_frame_trajviz; confirmed pixel-accurate
realsense_native	stretch_motion without artifacts (182)	Best-effort: coordinates divided by 1280×720 (the RealSense D435i native capture resolution per the collection script). No source ground truth is available for this subset.

---

Known Data Quality Issues

The following issues were identified during import and are preserved in the data:

1. Incomplete stretch_motion subset (182 groups): These groups have no pre-rendered visualizations (video_trajviz, last_frame_trajviz, opticalflow, storyboards are all absent). Only video_raw is available. These samples cannot be used with the paper's VLM evaluation methodology without regenerating the visualizations. Identified by has_source_artifacts == False or the "incomplete" tag.

2. num_steps vs trajectory_length disagreement (~160 rows): The source's num_steps field reflects the original capture length before some post-processing trimmed the trajectory. trajectory_length (= len(trajectory)) is the reliable count and is used for all frame-level annotations.

3. Trajectory partially outside frame: Some trajectories extend into negative coordinates or past the video edges. FiftyOne clips these gracefully at the frame border; no values are modified.

4. Variable video resolutions: Human demos span 14 different square resolutions (208×208 to 480×480 plus one 640×480). Stretch demos with artifacts use three resolutions (320×320, 352×352, 480×480). The incomplete stretch subset uses eight different resolutions (192×192 to 720×720).

---

Citation

@article{hwang2024motif,
  title={MotIF: Motion Instruction Fine-tuning},
  author={Hwang, Minyoung and Hejna, Joey and Sadigh, Dorsa and Bisk, Yonatan},
  journal={arXiv preprint arXiv:2409.10683},
  year={2024}
}

APA: Hwang, M., Hejna, J., Sadigh, D., & Bisk, Y. (2024). MotIF: Motion Instruction Fine-tuning. arXiv preprint arXiv:2409.10683.