zmq232/nymeriaplus-8seq-subset

NymeriaPlus — 8-sequence third-person (observer) subset

A private, research-only subset of NymeriaPlus (1,100 sequences total, ~67 TB), prepared for pose-to-video training where the observer RGB video is the prediction target. This repo contains 8 complete sequences: 8/1,100 of the dataset, chosen one-per-person across 2023-06-07 .. 2023-06-13 to cover different participants, homes and activity scripts.

⚠️ License. The data is governed by the NymeriaPlus Dataset License (Meta Non-Commercial Research License + CC BY-NC 4.0). It is for non-commercial research only. Section 3.1(c) prohibits hosting or transferring the content to any third party, which is why this repo is private and must stay private. Do not make it public, do not redistribute. This repo is not affiliated with or endorsed by Meta.

Contents

One zip per sequence under archives/, each expanding to <sequence_name>/… with the official nymeriaplus layout:

group	files	purpose
recording_observer_data_data_vrs	recording_observer/data/data.vrs	observer RGB video — the pose-to-video target
body_processed	body/xdata_smpl_neutral.npz, body/xdata_mhr.glb	per-frame SMPL + MHR human mesh
body_raw	body/xdata.{healthcheck,mvnx,npz}	raw XSens motion + health check (debugging)
slam	recording_{head,lwrist,rwrist,observer}/mps/slam/{closed,open}_loop_trajectory.csv, online_calibration.jsonl, semidense_points.csv.gz, summary.json	camera trajectories, online calibration, semi-dense points
timesync_and_imu	recording_{head,lwrist,rwrist,observer}/data/motion.vrs	time sync + IMU
narration	narration/{motion_narration,atomic_action,activity_summarization}.csv	motion captions and clip filtering
metadata_json	metadata.json	participant / scene / available-modality info
LICENSE	LICENSE	dataset license

Deliberately excluded (not downloaded): recording_head_data_data_vrs (first-person video, only needed for ego/exo pairing), recording_lwrist_* and recording_rwrist_* VRS, slam_semidense_observations (11.6 TB), audio, eye_tracking, object_bounding_box, object_mesh, video_main_rgb previews.

Sequences

sequence	date	participant	raw size
20230607_s0_james_johnson_act0_e72nhq	2023-06-07	james johnson	~11 GB
20230607_s1_barbara_wheeler_act0_cvxtsi	2023-06-07	barbara wheeler	~12 GB
20230608_s0_shelby_arroyo_act1_w436lf	2023-06-08	shelby arroyo	~17 GB
20230609_s0_angela_harrell_act2_42epvf	2023-06-09	angela harrell	~12 GB
20230609_s1_heather_becker_act0_pxfo7n	2023-06-09	heather becker	~12 GB
20230612_s0_jason_smith_act0_perjr7	2023-06-12	jason smith	~14 GB
20230612_s1_christina_jones_act0_u2r0z8	2023-06-12	christina jones	~12 GB
20230613_s0_virginia_rivera_act0_b6hopn	2023-06-13	virginia rivera	~13 GB

manifest.json records the exact archive sizes; every artifact was verified against the SHA-1 published in the official signed-URL JSON at download time.

20230613_s0_virginia_rivera_act0_b6hopn is shipped as 11 chunks under archives/parts/ instead of a single zip (the uplink to HF kept dying at ~2 GiB per connection). Rebuild it with reassemble.sh; the concatenation is byte-identical to the original archive (sha256 0f1bbbc26d9faf1663dfd0a943c9071991a8a22d470af6f3f88aede6a95afe52).

Usage

mkdir -p nymeriaplus_subset && cd nymeriaplus_subset
for z in ../archives/*.zip; do unzip -q "$z"; done

# only needed for the chunked sequence:
bash ../reassemble.sh ../archives/parts
unzip -q 20230613_s0_virginia_rivera_act0_b6hopn.zip

Then read it with the official tooling:

git clone https://github.com/facebookresearch/nymeria_dataset
cd nymeria_dataset && uv sync
uv run nymeriaplus-viewer -i /path/to/nymeriaplus_subset/20230607_s0_james_johnson_act0_e72nhq

Citation

@misc{nymeriaplus26,
  title={NymeriaPlus: Enriching Nymeria Dataset with Additional Annotations and Data},
  author={Daniel DeTone and Federica Bogo and Eric-Tuan Le and Duncan Frost and Julian Straub and Yawar Siddiqui and Yuting Ye and Jakob Engel and Richard Newcombe and Lingni Ma},
  year={2026}, eprint={2603.18496}, archivePrefix={arXiv}, primaryClass={cs.CV},
  url={https://arxiv.org/abs/2603.18496}
}