The deployment backend for physical AI

Turn fleet hours into compounding autonomy.

DataCore ingests robotics data over bad networks, keeps multi-sensor semantics intact, and streams synchronized slices for debugging and training.

Request a pilot See how DataCore works

We’re onboarding a small number of design partners for 2026 pilots.

Compatible with ROS / MCAP / bags Deploy managed cloud, VPC, or hybrid

Overview

Ingest → store → retrieve → operate

Robotics teams don’t need another dashboard. They need a backend that makes data capture reliable, retrieval fast, and operations auditable, so teams can ship improvements faster and with fewer regressions.

Ingest

Buffer locally, resume uploads across disconnects, and get durable acknowledgements.

Store

Append-only recordings with searchable metadata, clear retention, and project isolation.

Retrieve

Query time windows and stream synchronized slices (camera + LiDAR + state) without downloading entire logs.

Operate

Tie incidents, slices, and (when needed) teleop sessions back to the same recordings and audit trail.

DataCore is a deployment backend for robotics logs: it makes capture reliable, keeps semantics portable, and makes time-aligned retrieval a primitive.

  • Not just storage or a labeling UI
  • Not a replacement autonomy stack (yet 👀)

The problem

Robots don’t fail on intelligence first. They fail on the loop.

Outside the lab, autonomy breaks for boring reasons: the network drops, logs are incomplete, sensors drift out of sync, and nobody can replay what happened two weeks later.

Most teams hit the same bottlenecks:

  • The “important” moments don’t make it back from the field.
  • Multi-sensor logs aren’t reliably time-aligned, so debugging turns into guesswork.
  • Dataset creation becomes a pile of scripts nobody trusts, and nobody can reproduce.
Read our thesis (deployment gap)

Workflows

The workflows we’re built around

The goal isn’t “more data.” It’s less time between an incident and a fix.

Incident → slice → reproduce

Workflow A

A robot flags an event. You locate the recording session. You stream a synchronized slice around the time window. You replay it in your toolchain and ship a fix with confidence.

Outcome: debugging becomes a repeatable workflow, not a scavenger hunt.

Raw logs → dataset version (with provenance)

Workflow B

You define a selection rule (robots, environments, labels). DataCore generates a dataset version with clear provenance. You run a curation step (alignment, decoding, QA). You export to training and eval.

Outcome: you can explain performance changes with traceable data.

These are the first workflows we support. Over time, DataCore will expand toward the full robotics dev ops loop.

The architecture

DataCore is designed around two truths: networks are unreliable, and robotics data is multi-modal.

Edge

  • Capture
  • Buffer + upload
  • Teleop client

Cloud (DataCore)

  • EdgeRelay
  • TypeAtlas
  • Storage + index
  • Retrieval + pipelines

Tooling

  • Console
  • APIs + SDKs
  • Operator tools
Start walkthrough Stop walkthrough
EDGE Robot / device
CLOUD Platform planes
Control & Governance Identity • RBAC • Audit
TOOLING Console + APIs

01 / 10

EDGE

Capture

Capture synchronized multi-modal streams on the robot (e.g., mark a 20s window around a near-miss).

    Scroll to explore

    EdgeRelay

    Ingestion reliability

    Resumable ingestion built for unstable connectivity (durable acks, buffering, and backpressure).

    TypeAtlas

    Semantics that stay portable

    A semantics layer that keeps types, schemas, and deterministic transforms portable across ecosystems.

    Console

    UI you can operate from

    A daily workflow surface to browse recordings, request slices, and export artifacts into your toolchain.

    Teleoperation primitives

    Safe operator handoffs

    Safe-by-default operator handoffs with scoped permissions, guardrails, and an auditable trail tied to recordings.

    Security

    Security and deployment options

    Most robotics teams need flexibility more than slogans. We aim for least-privilege access, clear audit trails, and deployment options that match real constraints.

    • Projects are isolated by default
    • Access is scoped (RBAC + API keys + device identity)
    • Actions and artifacts link back to recordings
    • Retention and deletion are explicit

    Deployment

    Deployment: managed cloud, customer VPC/hybrid, or on-prem components when connectivity or regulation requires it.

    If you have strict requirements (data residency, retention, audit), tell us early. We’ll scope the pilot accordingly.

    Roadmap

    Pilot-ready DataCore (2026)

    • Production-grade ingestion (buffering, throttling, identity, enrollment)
    • Recording sessions and manifests and provenance
    • Retrieval v0: time-window queries and synchronized multi-stream slices
    • Export v0 into common robotics/ML workflows
    • Console MVP: org/projects and recording browse and slice retrieval

    Later in 2026

    • Dataset catalog with versioning, lineage, and reproducible pipeline runs
    • Incident-debug workflow, fleet observability primitives, and higher yield per fleet hour (QA + episode extraction)
    • Teleoperation support (MVP) and console expansion for datasets, exports, and pipelines

    Is DataCore a fit?

    Fast self-qualification for pilots.

    Good fit if you…

    • Operate outside controlled networks
    • Rely on multi-sensor, time-aligned logs
    • Spend too much time on log plumbing and dataset glue
    • Need fast retrieval for debugging and training (not just storage)
    • Want governance as teams and fleets scale

    Probably not a fit if you only need…

    • A labeling UI or a single-machine store for a prototype

    Request a pilot

    Request a pilot

    Leave an email. If it’s a fit, we’ll reply with a suggested pilot plan. No spam.

    Video LiDAR IMU Radar Audio

    If it’s a fit, we’ll reply with a suggested pilot plan. No spam.