Sky AI LAB Logo SKY AI LAB
Home Overview Custom Services & Workflows 10,000 Sq Ft Facility Data Processing & QA Pipeline Campaign Scope Calculator Scope Campaign →
In-House Processing Team 50+ Workstations • Multi-Pass Quality Assurance

From Raw Multi-Sensor Streams to
Zero-Defect ML Datasets

Data collection is only half the battle. Our dedicated team of 50+ data engineers and QA specialists inspects, time-synchronizes, transcribes, annotates, and formats every single episode for direct ingestion into your model training loop.

Configure Dataset Delivery → See QA Workflow ↓
END-TO-END DATA REFINERY

The 5-Stage Processing Engine

How our data team turns multi-camera feeds and human demonstrations into production-grade training tokens:

Stage 01 • Ingestion

PTP Genlock Synchronization

Raw 4K video feeds, EgoCAM streams, audio recordings, depth frames, and robot joint states are ingested over 100GbE optical fabric. Every frame is aligned to the IEEE 1588v2 hardware PTP clock with <0.08ms measured shutter jitter.

SYNC TARGET: ±0.038ms
DROPPED FRAMES: 0.00% Guaranteed
Stage 02 • Audio & Narration

Narration Audio Transcription

Spoken step-by-step human narrations are transcribed with high-accuracy speech models and audited by human verifiers. Each sentence, phrase, and sub-goal is given millisecond-accurate start/stop timestamps mapped to the exact action frame.

AUDIO CODING: 48kHz 24-Bit FLAC
VERIFICATION: 100% Human Audited
Stage 03 • Spatial Annotations

Action Chunking & 3D Annotations

Episodes are segmented into discrete semantic primitives (e.g. reach_carton, grasp_tape, fold_flap). Optional 21-keypoint 3D MANO hand tracking, 3D bounding boxes, and object masks are synthesized.

CHUNK HORIZON: k=16 / k=32 timesteps
3D KEYPOINTS: Sub-millimeter Ground Truth
Stage 04 • Multi-Pass QA

Triple-Gate Quality Assurance

Every recorded episode passes three independent gates before approval: 1. Automated kinematic limit & frame-continuity checks. 2. Senior data annotator visual audit for occlusion and task fidelity. 3. Engineering Director random sampling sign-off.

REJECTION THRESHOLD: Strict Zero-Tolerance
DEFECT RATE: <0.2% on final delivery
Stage 05 • ML Formatting

Format Conversion & Packaging

Datasets are compiled into target schemas: Open X-Embodiment RLDS (TFDS format), standardized HDF5 with LZ4 compression, Hugging Face lerobot format, or custom client WebDataset tarballs with PyTorch/JAX DataLoaders.

SCHEMAS: RLDS / HDF5 / LeRobot
COMPATIBILITY: PyTorch & JAX Ready
Security & Air-Gap

Air-Gapped Optical Egress

Client data is isolated on private VLANs with physical encryption at rest (AES-256). Datasets are dispatched via private AWS S3 buckets, Snowflake share, or physical encrypted enterprise SSD drop-ships under strict MNDA.

COMPLIANCE: SOC 2 Type II Audited
DATA OWNERSHIP: 100% Client Exclusive
PRODUCTION READY CODE

Native Compatibility With Frontier Frameworks

Plug our datasets directly into OpenVLA, π0, Octo, ACT, and Diffusion Policy pipelines:

import torch
from torch.utils.data import Dataset, DataLoader
import h5py

class SkyAiMultimodalDataset(Dataset):
    """
    Sky AI LAB — Multimodal Human Demonstration & Narration Loader
    Time-synchronized 4K EgoCam + Audio Tokens + Bimanual Joint Trajectory
    """
    def __init__(self, h5_path: str, chunk_size: int = 16):
        self.f = h5py.File(h5_path, 'r')
        self.chunk_size = chunk_size
        self.images = self.f['/observations/images/ego_head']     # (N, 3, 224, 224)
        self.narration = self.f['/language_instruction'][:]      # Text instruction tokens
        self.actions = self.f['/action'][:]                      # (N, 14) Bimanual actions

    def __len__(self):
        return len(self.actions) - self.chunk_size

    def __getitem__(self, i):
        return {
            'image': torch.tensor(self.images[i] / 255.0, dtype=torch.float32),
            'action_chunk': torch.tensor(self.actions[i:i+self.chunk_size], dtype=torch.float32),
            'instruction': str(self.narration[i])
        }

# Batch loading ready for OpenVLA / ACT imitation policy fine-tuning
loader = DataLoader(SkyAiMultimodalDataset('skyai_packaging_task_088.h5'), batch_size=32, shuffle=True)
Scope Your Dataset Specifications →