Product
ACTUATE
Raw data in.
Training datasets out.
Our end-to-end processing pipeline takes raw, multi-modal sensor streams and transforms them into clean, structured, training-ready datasets — with every step automated, auditable, and optimised for robotics foundation model training.
Only clean data enters the pipeline.
Every raw sequence passes automated frame-level and sequence-level quality checks before any downstream processing. The QA stage flags and rejects blur, sensor dropout, occlusion events, motion artifacts, and synchronization failures — ensuring only high-integrity data moves forward. Each accepted episode receives an Episode Integrity Score (EIS): a transparent, weighted composite certificate across all check dimensions, not a binary pass/fail.
What happens here
- Frame-level blur & sharpness scoring
- Sensor dropout & gap detection
- Motion artifact flagging
- Sequence completeness & sync-drift validation
- EIS — Episode Integrity Score composite certificate
Every sensor stream, aligned to ±2ms.
Hardware-level temporal synchronization across all modalities: egocentric RGB, depth, IMU, and 3D pose. We use timestamped sensor buses and a proprietary alignment algorithm to achieve ±2ms precision — so every frame of video corresponds exactly to the correct IMU sample and pose keyframe. IMU-to-video interpolation is applied where hardware jitter would otherwise introduce drift.
What happens here
- ±2ms cross-modality alignment
- Hardware timestamp reconciliation
- IMU-to-video interpolation
- Drift correction across long sequences
- Verified sync metadata per sequence
Language grounded in real video frames.
Natural-language task instructions are generated by a vision-language model sampling actual frames from each episode — not filled-in sentence templates. The result is genuinely varied, per-episode language annotation grounded in what the model sees: object identities, scene state, and action context. Semantic parsing then adds per-frame and per-segment labels aligned to a unified action ontology covering task boundaries, object identity, grasp type, contact state, and environment context.
What happens here
- VLM-grounded language annotation (not templates)
- Per-episode natural-language instructions from sampled frames
- Per-frame semantic action labels
- Task boundary segmentation
- Object identity & state tracking
- Contact & grasp type annotation
Raw motion into structured action primitives.
Maps raw kinematic demonstrations to structured action primitives — grasp, place, push, carry, rotate, press — with full joint angle trajectories and contact state annotations. Every detected primitive carries a per-detection confidence score, not a boolean label. Confidence propagates through to episode-level and task-level aggregates, giving downstream consumers a transparent signal about labeling certainty at every granularity. Output is compatible with ACT and Diffusion Policy architectures.
What happens here
- Action primitive decomposition
- Per-detection confidence scores (not boolean labels)
- Episode & task-level confidence propagation
- Joint angle trajectory export
- Contact state annotations
- ACT / Diffusion Policy compatible
Training-ready datasets in your format.
The final stage packages all processed data — synchronized streams, annotations, action labels, and metadata — into your required output format. We support HDF5, RLDS, LeRobot, and custom schemas. For customers building embodied or VLA policies, the package includes metric-3D depth maps and retargeting-eligibility flags per episode, enabling direct use in sim-to-real transfer and cross-embodiment retargeting workflows. Every dataset ships with versioning, train/val/test splits, and a full metadata manifest.
What happens here
- HDF5, RLDS, LeRobot formats
- Custom schema support
- Metric-3D depth maps
- Retargeting-eligibility flagging per episode
- Train / val / test splits
- Full metadata manifest & versioning
Certification
Every episode,
certified.
Customers don't just receive labeled data — they receive a transparent quality trail per episode. Each delivered dataset carries a structured certificate covering integrity, labeling confidence, and retargeting readiness, so teams can make informed decisions about what goes into training.
EPISODE INTEGRITY SCORE
A transparent, weighted composite across all QA dimensions — blur, coverage, sync-drift, frame-level checks — delivered as a named certificate per episode, not a hidden internal metric.
CONFIDENCE BREAKDOWN
Per-detection confidence scores on every action primitive, aggregated to episode and task level. Customers can inspect labeling certainty at any granularity before training.
RETARGETING ELIGIBILITY
Each episode is flagged for cross-embodiment retargeting readiness based on metric-3D depth quality and kinematic coverage — relevant for teams building embodied or VLA policies.
AUDITABLE QUALITY TRAIL
Every delivered dataset includes a machine-readable metadata manifest recording QA outcomes, sync verification, annotation provenance, and labeling confidence — per episode.
Output formats
Delivered in
your format.
HDF5
Standard hierarchical format, universal compatibility
RLDS
Google's Robot Learning Dataset Spec
LeRobot
Hugging Face LeRobot-compatible format
Custom
Any schema your pipeline requires