API Documentation

What it does

The haptal API gives you programmatic access to pre-processed robotics datasets — compressed 450x from raw sensor data into dense embeddings that work across sensor types (GelSight, DIGIT, BioTac, custom). No collection rigs, no preprocessing, no sensor-specific code.

Why use it

• Skip infrastructure — A tactile data rig costs $50K+ and months. One API call replaces it.
• Sensor-agnostic — Embeddings normalize across hardware. Your model generalizes without retraining per sensor.
• Multiple formats — NumPy, PyTorch, TensorFlow, HDF5. Stream in real-time or batch download.
• Version pinning — Lock dataset versions for reproducible experiments across your team.

Install

pip install haptal-sdk

# No install needed — use curl directly

npm install @haptal/sdk

Authenticate

from haptal import HaptalClient

client = HaptalClient(api_key="your_api_key_here")
print(client.ping().ok)  # True

export HAPTAL_API_KEY="your_api_key_here"

curl -H "Authorization: Bearer $HAPTAL_API_KEY" \
  https://api.haptal.ai/v1/ping

const { HaptalClient } = require('@haptal/sdk');
const client = new HaptalClient({ apiKey: 'your_api_key_here' });
console.log((await client.ping()).ok);  // true

Fetch a dataset

data = client.datasets.fetch(
    dataset_id="humanoid-manipulation-v2",
    split="train",
    format="numpy"
)

print(data.observations.shape)  # (50000, 7, 128)
print(data.labels.unique())     # ['grasp', 'lift', 'place', ...]

curl -H "Authorization: Bearer $HAPTAL_API_KEY" \
  -X POST "https://api.haptal.ai/v1/datasets/humanoid-manipulation-v2/download" \
  -H "Content-Type: application/json" \
  -d '{"split":"train","format":"numpy"}'

const data = await client.datasets.fetch({
  datasetId: 'humanoid-manipulation-v2',
  split: 'train',
  format: 'numpy'
});
console.log(data.observations.shape);  // [50000, 7, 128]

Stream into training loop

stream = client.streams.connect(
    dataset_id="humanoid-manipulation-v2",
    batch_size=64,
    shuffle=True
)

for batch in stream:
    loss = model.train_step(batch.observations, batch.actions, batch.rewards)

Home · founders@haptal.ai