This tutorial runs one real recording through HFlow's recommended starting point for egocentric video: the default deterministic quality checks plus two hosted model checks. It demonstrates the useful split in one command:
- HFlow measures recording integrity locally and deterministically.
- HFlow's hosted checks add semantic evidence without a model server of your own. Contact us to get access to the hosted API.
- Both kinds of evidence are versioned and recorded through the same pipeline.
The hosted checks are registered explicitly rather than included in
hflow.checks.DEFAULT_CHECKS. Constructing a normal hflow.App must remain
offline and must not send recording data over the network unexpectedly. This
tutorial makes that boundary visible in both the command and the source.
From the repository root:
uv sync --locked
uv tool install -U huggingface_hubThe sample is a short first-person "Thimble Removal" episode from Lightwheel's
EgoDemo dataset. The
Hugging Face CLI returns its cache path; copy it to the short path used below:
mkdir -p data/episode-evaluation
downloaded_sample_mcap_path="$(
hf download LightwheelAI/EgoDemo \
'EgoStand/mcap/Thimble Removal/a8d29fea-3cf9-47f7-ad4b-4b1d0ecb7a71.mcap' \
--repo-type dataset \
--quiet
)"
cp "$downloaded_sample_mcap_path" data/episode-evaluation/sample.mcapIf Hugging Face requests access, accept the dataset's terms and run
hf auth login, then repeat the download.
The sample has two camera streams, so select its left head camera explicitly:
uv run python examples/evaluate_episode.py \
data/episode-evaluation/sample.mcap \
--hosted-base-url "$HFLOW_HOSTED_BASE_URL" \
--camera /sensor/camera/head_left/videoSet HFLOW_HOSTED_BASE_URL to the hosted API URL you received. HFlow
processes the episode locally; the two model checks send only the selected
JPEG frame to that URL.
The first sample frame visibly contains two wearer hands manipulating material. The hosted results should therefore include:
build_ai/hand_count/prediction = 2
build_ai/active_manipulation/prediction = yes
The report also includes HFlow's automatic deterministic baseline:
| Check | Evidence it records |
|---|---|
episode_duration |
Recording duration and message coverage |
timestamp_regularity |
Timing intervals, gaps, and cross-stream alignment |
camera_frame_stats |
Frame count, blackout, freeze, and exposure measurements |
keyframe_interval |
Video seekability evidence |
content_digest |
Exact recording identity for duplicate detection |
media_digest |
Visual-media identity across differently wrapped recordings |
build_ai_hand_visibility |
Number of wearer hands visible in the selected frame |
build_ai_active_manipulation |
Whether the wearer is visibly manipulating an object |
The canonical MCAP, contact sheets, run evidence, and Parquet catalog are
written beneath data/episode-evaluation/.
Replace the sample path with your MCAP. Omit --camera when it contains only
one camera, or pass the desired camera topic when it contains several:
uv run python examples/evaluate_episode.py path/to/episode.mcap \
--hosted-base-url "$HFLOW_HOSTED_BASE_URL" \
--camera /your/egocentric/cameraUse --frame-time-seconds to choose another instant. The selected time and
camera are part of each model check's version, so results from different
sampling configurations do not silently claim to be comparable.
The hosted checks implement Build AI's published hand-visibility and active-manipulation prompts. For custom prompts, models, or OpenAI-compatible endpoints, continue with the Build AI evaluation guide. To configure or add deterministic checks, see the built-in checks guide.
This starter pipeline is intentionally general. Real deployments usually need checks and processing tailored to the recording rig, task, failure modes, and training format—and a plan for running them across the full corpus.
If you want help designing that pipeline or evaluating what HFlow can automate for your team, tell Hebbian Robotics about your workflow. For implementation questions and open-source feedback, join the HFlow Discord.