Mission / Investor thesis

Maximize the time physical AI companies spend validating customer value.

Reduce device integration and failure investigation time. Shorten the path back to the next valuable real-world experiment.

Investor thesis

Turn debugging time into customer-value validation time.

RFR shortens the time from a physical AI failure to the next valuable experiment.

The initial wedge is a Connectivity Failure Reconstruction Sprint.

Adapters, Failure Patterns, and capture requirements learned through integration, detection, and reconstruction become reusable product knowledge.

Compounding loop

Every failure solved adds knowledge that shortens the next investigation.

Start by reconstructing one resolved connectivity failure from three to five data sources.

Register validated modem notifications, normal connection sequences, reconnect patterns, and Failure Patterns.

Repeated observation gaps become Ingest Agent, BLE Node, Edge Gateway, and ring-buffer product requirements.

Connectivity SprintOne resolved failure becomes a Device-to-Cloud Timeline
NormalizeVendor logs map to a Common Connectivity Model
Evidence ChainSeparate deviation, evidence, uncertainty, and missing data
Adapter RegistryRegister modem events, firmware differences, and normal/failure patterns
Distributed CaptureDeploy BLE-only Nodes and Gateway where evidence is repeatedly missing
Team WorkflowReuse Failure Clips for review, comparison, and regression

Why now

Robots are entering real operations and becoming distributed AI systems.

4.66M unitsIndustrial robots operating worldwide in 2024, up 9% year over yearIFR World Robotics 2025
~200K unitsProfessional service robots sold in 2024, up 9% year over yearIFR Service Robots 2025

As robots move into operations, time intended for validating customer value is absorbed by integration and investigation.

AI, control software, hardware, connectivity, and cloud services interact, so every local component can look healthy while the end-to-end task fails.

When failure logs become reusable Evidence, the same class of failure can be resolved faster next time.

RFR is a physical AI development foundation focused on Time to Next Valuable Experiment.

Pain

Logs exist, but the device-to-cloud sequence is missing.

A task can fail in LTE communication even when motors and sensors are healthy.

Application, Linux, modem, radio, transport, and cloud logs use different clocks, formats, and owners.

Successful retries and reconnects bury the first deviation under recovery activity.

RFR returns how far the request traveled, where it first deviated, and what evidence remains missing.

Wedge

Connectivity Failure Reconstruction Sprint

Reconstruct one resolved mobile-connectivity failure from three to five data sources.

Deliver a Device-to-Cloud Failure Timeline, First Observed Deviation, Evidence Chain, failure scope, uncertainty, Missing Evidence, next-capture plan, and recurrence-rule candidates.

Device integration and wireless readiness are a parallel entry point through a Wireless Readiness Sprint.

Initial customers

Robotics teams connecting machines to cloud services over mobile networks.

  • Roughly 5–30 people building hardware and software in-house
  • Logs distributed across ROS 2, applications, Linux, modem, and cloud
  • Long investigations of LTE timeout, reconnect, and task-resume failures
  • No access to mobile network operator internals, but a need to narrow the scope
  • No dedicated observability team and repeated movement across tools

Expansion

From reconstruction service to a Device-to-Cloud Evidence Layer.

RFR Ingest Agent captures ROS 2, Linux, modem, network, cloud communication, and ring buffers.

Common Connectivity Model and Adapter Registry turn modem and firmware differences into reusable knowledge.

The product expands into BLE-only Nodes, Edge Gateway, Failure Patterns, regression testing, and Team Workflow.

Defensibility

Cross-layer Evidence, Connectivity Adapters, and Failure Patterns compound together.

  • Device-to-Cloud Timeline across application, Linux, modem, LTE, transport, and cloud
  • Connectivity Adapters for modem AT notifications, firmware differences, and connection states
  • Evidence Model separating First Observed Deviation, uncertainty, and Missing Evidence
  • Registry of normal connection sequences, reconnect patterns, Failure Patterns, and regression conditions
  • Distributed Capture separating BLE Mesh control plane from Direct BLE data plane
  • Raw-first operation where AI candidates become deterministic only after validation
InitialConnectivity Reconstruction + Manual alignment
Mid-termConnectivity Adapters + Ingest Agent + Pattern Registry
Long-termDistributed Capture + Team Workflow + Regression

Business model

Move from fixed-scope reconstruction to a recurring Failure Evidence product.

Initial ServiceConnectivity Failure Reconstruction Sprint: one failure, three to five data sources
Parallel EntryWireless Readiness Sprint: observe, translate, and stress-test one existing device
DeliverableDevice-to-Cloud Evidence Bundle and recurrence-rule candidates
Recurring ProductAdapter Registry, Failure Patterns, Team Workflow, and regression testing
Capture ExpansionIngest Agent, BLE-only Node, Direct BLE retrieval, and Edge Gateway

What we prove first

Can one LTE failure be scoped faster and more reproducibly?

  • Reconstruct a Device-to-Cloud Timeline from three to five logs
  • Separate the first deviation from later recovery activity
  • Make unavailable operator-internal evidence explicit
  • Enable a second engineer to reach the same evidence and scope
  • Convert the result into a recurrence-detection rule

Measurement

Compare RFR with the team's existing workflow, including Foxglove.

Use a resolved historical failure and an engineer who was not involved in the original incident.

  • A: Customer's original workflow
  • B: Foxglove + terminal + existing scripts
  • C: RFR Timeline
  • D: RFR Timeline + AI summary
  • Measure failure-layer hypothesis, evidence events, wrong hypotheses, tool moves, and alignment time

Reproducibility between two analysts matters as much as speed.

Validation metrics

Measure Time to Next Valuable Experiment.

  • Time to First Evidence
  • Time to Failure Scope
  • Time to Next Experiment
  • Tool Handoffs
  • Engineer Hours per Incident
  • Adapter Reuse Rate
  • Value Validation Ratio

Competitive position

Connect existing tools instead of replacing them.

Logic analyzers, terminals, RTT/JTAG, and ROS/MCAP tools are strong at their own layer.

Physical AI failures require the causal sequence from intent through physical response and cloud interaction.

RFR ingests failure triggers and low-level events, preserves raw Evidence, and exports JSONL or MCAP when needed.

Logic analyzerStrong for signals, but rarely connected to AI intent or robot state.
Serial terminalUseful for UART logs, but weak across multiple nodes, Failure Clips, and grounded AI summaries.
RTT / JTAGStrong inside an MCU, but not designed for a running robot's end-to-end failure timeline.
ROS / MCAP toolsStrong for visualization and playback. RFR connects through MCAP import and export.

Productization sequence

Solve real failures, then productize validated knowledge and recurring observation gaps.

Step 1

Connectivity Sprint

Reconstruct one resolved LTE failure from three to five logs.

Step 2

Connectivity Model

Normalize Application, Modem, Registration, Session, Transport, and Recovery.

Step 3

Adapter Registry

Register modem notifications, firmware differences, normal sequences, and Failure Patterns.

Step 4

Ingest Agent

Continuously capture Linux, modem, network, cloud communication, and ring-buffer state.

Step 5

Distributed Capture

Notify over BLE Mesh and retrieve selected Failure Clips over Direct BLE.

Step 6

Team Product

Expand into Failure review, normal/failure comparison, recurrence detection, and regression testing.

Company / Builder

Built by invilab from hands-on product and field-debugging work.

invilab designs technology and creates new value. Its vision is to contribute to a future that makes people happier, and its mission is to design product development that accelerates innovation and value creation.

RFR emerged from that product-development and technical-design practice as a Device-to-Cloud Failure Reconstruction PoC for physical AI.

Founder

Ryota Wakabayashi

Representative Director, CEO & CTO / Co-founder

Studied applied physics at Waseda University and its graduate school, researching quantum communication. After working in R&D at Sony Corporation, he co-founded invilab.

Company

invilab Inc.

Founded May 7, 2019

Capital: JPY 5 million

3F Koinuma Building, 1-10-7 Kusunokidai, Tokorozawa, Saitama, Japan

contact@invilab.co.jp

View invilab company information