robotics annotation infrastructure

The data layer robots learn from

Annotation and behavior failure tagging for robot training pipelines

accuracy

faster

cheaper

3D LiDAR· Egocentric· Teleoperation· Failure Taxonomy· Policy Filtering· LeRobot v2· RLDS· HDF5· Open X-Embodiment· grasp_slip· object_drop· overcorrect· BiLSTM· Quality Score· 3D LiDAR· Egocentric· Teleoperation· Failure Taxonomy· Policy Filtering· LeRobot v2· RLDS· HDF5· Open X-Embodiment· grasp_slip· object_drop· overcorrect· BiLSTM· Quality Score·

how it works

Raw data in. Training-ready out.

Ingest

Any format off your robot.

any format

Pre-label

Open source models run first. High accuracy before human review.

AI-first

Haptal Engine

Our proprietary core.

failure detection

quality scoring

policy filtering

anomaly modeling

Human review

Experts verify only what the model flags.

verified

Delivery

Your format. Failure tags embedded. Plug in and train.

training ready

output

Your format. Zero conversion.

train.py

# zero changes to your training script

LeRobotDataset("haptal/your_robot")

# use_for_policy → true / false

# failure_tag → grasp_slip | object_drop | null

# quality_score → 0.0 – 1.0

clean = dataset.filter(lambda ep: ep["use_for_policy"])

train_policy(clean)

LeRobot v2 RLDS HDF5 Zarr Open X-Embodiment

Your robots are generating data right now

We will tell you which episodes are clean and which should never go into training

No spam. Early clients get one free annotated session.