bathrlor/DynamicVLA-DOM-Validation-Videos

DynamicVLA DOM Validation Videos

This dataset contains all 1,800 validation videos from the published DynamicVLA-IsaacSim45-PathTracing campaign. The run completed 90 official DOM configurations with 20 trials each.

Result

Item	Value
Completed trials	1800
Successes	664
Failures	1136
Observed success rate	36.89%
Tasks with zero successes	23
Tasks with 20/20 successes	4

The result is a controlled non-benchmark experiment. It is not the paper's native two-A6000 benchmark result.

Fixed protocol

Component	Value
DynamicVLA source	a27a06d2ca74d0e987a5e552e01013073e93cfd8
Checkpoint	hzxie/dynamic-vla-DOM
Checkpoint revision	c9609d94c5940eb42436a31927d7e80ae454279e
Isaac Sim	4.5.0.0
Isaac Lab source	2.2.1
Renderer	Path Tracing
Control rate	25 Hz
Physics and control step	40 ms
Virtual inference latency	200 ms
Emulated chunk age	5 control steps
Single GPU emulates dual GPU	yes
benchmark_eligible	false

The simulation clock is frozen while CUDA inference runs. A new action chunk becomes available after the virtual 200 ms deadline, allowing the previous queue to advance five control steps.

Files

metadata/
  campaign.json
  episodes.csv
  episodes.jsonl
  results.json
  summary.csv
  progress.json
checksums/
  metadata.sha256
  videos.sha256
videos/
  <task>/trialNNN__SUCCESS.mp4
  <task>/trialNNN__FAIL.mp4

Hugging Face is the only video host. The videos/ tree supports task-level selective download, and the GitHub repository is the guide for installation and reproduction:

https://huggingface.co/datasets/bathrlor/DynamicVLA-DOM-Validation-Videos

https://github.com/tanboning1118/DynamicVLA-IsaacSim45-PathTracing

metadata/episodes.csv is the authoritative trial index. A video filename, task identifier and success label can be joined to trial status, video path, save time, frame count, FPS and evaluation variant through that file.

Verify a download

Download one capability group and verify the individual files:

hf download bathrlor/DynamicVLA-DOM-Validation-Videos \
  --repo-type dataset --local-dir DynamicVLA-DOM-Validation-Videos \
  --include 'videos/1-1_*/*'
cd DynamicVLA-DOM-Validation-Videos
sha256sum -c checksums/videos.sha256

Check the trial accounting:

python - <<'PY'
import csv
with open("metadata/episodes.csv", newline="", encoding="utf-8") as f:
    rows = list(csv.DictReader(f))
print("trials", len(rows))
print("successes", sum(r["status"] == "SUCCESS" for r in rows))
PY

Expected result:

trials 1800
successes 664

Source and licensing

The evaluation harness is available at:

https://github.com/tanboning1118/DynamicVLA-IsaacSim45-PathTracing

The videos are generated evaluation artifacts. Model weights, USD scenes, DOM raw data and simulator binaries are not redistributed in this dataset. The upstream DynamicVLA S-Lab License 1.0, including its non-commercial restriction, applies to derived work. Isaac Sim, Isaac Lab and DOM assets have their own terms. This dataset does not imply author or NVIDIA endorsement.

Citation

@inproceedings{xie2026dynamicvla,
  title     = {{DynamicVLA:} A Vision-Language-Action Model for Dynamic Object Manipulation},
  author    = {Xie, Haozhe and Wen, Beichen and Zheng, Jiarui and Chen, Zhaoxi
               and Hong, Fangzhou and Diao, Haiwen and Liu, Ziwei},
  booktitle = {NeurIPS},
  year      = {2026}
}