Jinyan0924/qwen_robotics_open_dataset_egowalk

EgoWalk navigation frames

EgoWalk (MIT) converted into a per-frame training format for image-in, trajectory-out navigation policies: 57 hours and 232 km of people walking in Moscow, indoors (malls, stations, offices) and outdoors, filmed by a chest-mounted ZED camera, with visual odometry and about 79,000 language goals ("Walk to the glass door on the left").

It is one of the training sources of qwen_robotics_open_dataset, which builds training data usable by both Qwen-RobotNav-style and Qwen-VLA-style models from open datasets. Conversion code: scripts/convert_egowalk.py.

Top: the current frame and a prompt made at load time. Bottom: the same sample in the walker's frame: past positions, the 8 target waypoints taken every 0.25 m along the path, and the goal.

What is in it

Split	Recordings	Episodes	Frames	Hours	km	Indoor frames	Language segments
train	215	738	905,403	50.3	202.2	47%	69,143
validation	9	37	37,713	2.1	9.7	36%	2,881
test	15	51	89,863	5.0	19.8	43%	7,058

• Frames at 5 Hz, 960 × 600 RGB (JPEG, quality 90, re-encoded from the source video), every frame of the source with a valid pose. Depth video is left out: the policy sees RGB only.
• Poses: x, y, z, yaw in metres / radians in the episode's odometry frame (z up), straight from the source's ZED visual odometry, not resampled or normalised. Yaw is the camera's heading; it matches the walking direction to within a few degrees.
• Episodes: a recording is cut wherever odometry failed (missing poses, re-initialisation jumps over 1 m, gaps over 1 s); episodes shorter than 20 frames or 3 m are dropped. 826 episodes from 239 recordings.
• Language goals (episodes.segments): both annotation sets of the source, end2end (VLM-written) and goal_boxes (detector-based). A segment's instruction names what the walker reaches at end_frame.
• Indoor / outdoor: estimated per frame with CLIP ViT-L/14 zero-shot (frame_indoor_prob); an episode is indoor / outdoor if at least 80% of its frames are, else mixed. Spot-checked by eye, not annotated.
• Splits are by recording (hash of its name), 90 / 5 / 5.
• Walking speed: median 1.2 m/s, much faster than a 0.5 m/s robot. Speed is deliberately not baked into anything stored; targets are cut by distance along the path at load time.

The per-frame format

Two configs, joined by episode_id.

frames: one row per frame, rows of an episode contiguous and in time order.

Column	Type	Meaning
episode_id	string	egowalk/<recording>/<nn>
frame_index	int32	0-based within the episode
source_frame	int32	frame number in the source recording
timestamp	float64	seconds since the episode's first frame
image	image	front camera (left camera of the ZED)
image_right	image	right camera of a stereo pair; None here (the source publishes the left only)
pose	float64[4]	x, y, z (m), yaw (rad), episode odometry frame, z up

episodes: one row per episode.

Column	Meaning
dataset, source_sequence, split	origin
num_frames, rate_hz, duration_s, path_length_m, median_speed_mps	size and speed
embodiment	person_walking
environment, environment_method, indoor_fraction, frame_indoor_prob	indoor / outdoor estimate, per episode and per frame
camera	name, width, height, K (row-major 3 × 3), distortion_model (opencv_rational), distortion (k1 k2 p1 p2 k3 k4 k5 k6), height_m above the ground (1.20–1.36 m), stereo_baseline_m
segments	language goals: start_frame, end_frame, task, instruction, brief, source
ends_at_rest	the episode ends with the walker standing still (a real "stop here"); false at odometry cuts
fits	both (has language goals: usable for Qwen-RobotNav and Qwen-VLA style training) or robotnav
licence	source licence

Load and cut training samples

from datasets import load_dataset
from hnod.windows import FrameWindows, WindowConfig   # github.com/naomili0924/qwen_robotics_open_dataset

repo = "Jinyan0924/qwen_robotics_open_dataset_egowalk"
# name the split's files: load_dataset(repo, "frames", split="validation") downloads every split first
frames = load_dataset(repo, data_files={"validation": "data/frames/validation-*.parquet"}, split="validation")
episodes = load_dataset(repo, data_files={"validation": "data/episodes/validation-*.parquet"}, split="validation")
samples = FrameWindows(frames, episodes, WindowConfig(n_waypoints=8, spacing_m=0.25))
s = samples[0]
s["images"]   # 11 PIL images: the current frame and one every 0.5 s before it
s["prompt"]   # "Please walk towards the goal (7.9, 2.6)." or "Walk to the glass door. It is at (3.0, 1.5)."
s["target"]   # (8, 3) waypoints x, y, yaw in the current frame: x forward, y left, metres
s["stop"]     # (8,) true past the end of an episode that ends at rest

The train split is 76 GB, and datasets keeps both the downloaded files (HF_HOME) and its own Arrow copy (cache_dir): put both on a disk with about 160 GB free.

Samples are cut at load time, so horizon, waypoint spacing, history length and the share of language prompts are parameters, not properties of the data. Point goals lie 4–20 m further along the walked path (beyond the waypoint horizon, so the goal does not give away the answer); language goals are the end of a segment that is still at least the horizon ahead. Standing-still odometry jitter (under 5 cm) does not count as travel.

Limitations

• A person, not a robot. Camera at chest height (about 1.3 m), walking at about 1.2 m/s, through places and gaps a robot may not fit. No obstacle ground truth, so it is training data only, never evaluation data.
• Odometry, not ground truth. ZED visual odometry drifts and sometimes fails (hence the episode cuts).
• Wide-angle lens with strong distortion (rational model given in camera); images are not undistorted.
• Indoor / outdoor labels are estimates (CLIP), roughly right by spot check; mixed scenes (covered walkways, station entrances) are uncertain.
• Language goals are machine-written by the source authors (VLM and detector pipelines) and not verified.

Licence and citation

EgoWalk is MIT-licensed; this conversion keeps the MIT licence. Cite the EgoWalk authors (dataset, API).