Technology / Technical details

Put device communication, Linux, LTE, and cloud evidence on one timeline.

RFR normalizes available evidence into Semantic Events and reconstructs First Observed Deviation, Evidence Chain, uncertainty, and Missing Evidence in a Failure Timeline.

Product path / Vision and current scope

Start with Device-to-Cloud Failure Reconstruction, then productize the observation points and Adapters repeatedly required.

The home page explains the product vision. This page shows the implementation sequence.

Current capability and future architecture stay separate so safety and customer value can be validated at every stage.

Product VisionConnect device, ROS 2, Linux, modem, LTE, transport, cloud, and recovery behavior through shared semantics and time.
Current v0Three to five logs + Manual Anchor + Failure Clip + Evidence Chain.
Next PoCProductize recurring transformations, ring-buffer controls, BLE observation points, and recurrence-detection rules.
ComponentRole
RFR Ingest AgentRead-only ingestion of ROS 2, application, Linux, network, modem, cloud communication, and ring-buffer events.
RFR Tap / Bridge NodeObserve and normalize UART and RS-485 outside Linux. Operate as an active Bridge only when explicitly validated.
RFR Edge GatewayManage BLE Mesh, Direct BLE retrieval, clock coordination, and Failure Clip collection. Add Ethernet, Wi-Fi, or LTE only when required.
RFR Reconstruction CoreNormalize Semantic Events, map clocks, detect anomalies, construct Failure Timelines and Evidence Chains, and register patterns.

BLE architecture / Control plane and data plane

Every Node is BLE-only. Raw transfer and external connectivity are separate.

RFR does not stream raw logs continuously over BLE Mesh. The authoritative raw record stays on each Node, and only the required Failure Clip is retrieved over Direct BLE.

Wi-Fi and LTE are optional Edge Gateway capabilities. When LTE is present, it is also an observed target from modem state through cloud response.

BLE-only Node

Capture locally

Keep UART / RS-485 capture, ring buffers, timestamps, and drop counters locally. No Wi-Fi or LTE.

BLE Mesh

Control and notification

Node discovery, health, anomaly notification, configuration, time mapping, and clip-retrieval requests.

Direct BLE

Retrieve selected evidence

The Gateway connects directly to the target Node and retrieves only the event-index window, with retry and integrity checks.

Edge Gateway

Aggregate and analyze

Handle Direct BLE retrieval and local analysis. Add Wi-Fi or LTE only when cloud connectivity becomes necessary.

Failure Clip retrieval: Node detects an anomaly → raw stays local → BLE Mesh sends an anomaly notification → Gateway connects by Direct BLE → selected window is retrieved → RFR Core reconstructs the failure.

Common Connectivity Model

Normalize modem-specific notifications into Device-to-Cloud Evidence.

AI proposes Adapter candidates from data sheets and AT-command manuals.

After a human validates meaning and transformation quality, the Adapter runs deterministically without AI.

Connectivity events

  • Application request / timeout / retry
  • Modem ready / resetting / unavailable
  • Network searching / registered / lost
  • Data session activating / active / degraded
  • Signal strength / quality / cell / mode
  • DNS / TCP / TLS / MQTT / HTTP
  • Reconnect / modem reset / interface restart

Evidence rules

  • Raw events do not require trace_id at capture time
  • Attach transaction_id, clip_id, and derived trace after ingestion
  • Begin clock alignment with human Manual Anchors
  • Show uncertain ordering as uncertain
  • Do not turn unknown timestamp quality into a numeric claim
  • Keep operator-internal LTE state as Missing Evidence when unavailable

MVP scope / What comes first

v0 is a vertical PoC for one resolved mobile-connectivity failure.

The entry point is three to five existing data sources: ROS 2 or application logs, Linux, modem AT, LTE state, transport, cloud, or power.

Dedicated Taps, BLE Nodes, and Gateways come later, after existing logs reveal which observation points and Adapters are repeatedly missing.

Included in v0

  • Import MCAP, CSV, JSONL, journal, AT, and cloud access logs
  • Multi-lane ROS, Linux, Modem, Connectivity, Cloud, Power, and AI timeline
  • Manual Anchor plus lane offset and drift adjustment
  • Explicit unknown time and uncertain ordering
  • Failure Clip from range selection, operator marker, or deterministic anomaly
  • Raw Inspector, AI summary, and evidence-event jumps
  • Read-only tail, import, and ROS mirror

Not included in v0

  • A complete general-purpose observability platform
  • SDK-first integration
  • Mandatory trace_id on raw events
  • Fully automatic time sync for logs without a shared reference
  • Presenting commodity taps as production capture hardware
  • Mass-production design of a dedicated Capture Node
  • AI-generated control commands

Phase 0.5 / RX-only capture

Use commodity adapters to test whether low-level evidence improves diagnosis before building dedicated hardware.

The goal is not production-grade capture. It is same-day validation of whether low-level logs shorten failure isolation.

Adapter TX stays disconnected. Host-arrival time is never presented as a precision hardware timestamp.

UART tap / RX-only

Observe two TX lines with two USB-UART RX inputs.

  • Never connect Adapter TX
  • Check 1.8V / 3.3V / 5V levels
  • Never connect RS-232 levels directly to TTL UART
  • Share ground only when electrically safe
  • Do not use on safety-critical, high-voltage, or production equipment

RS-485 tap / Receive-only

Observe A/B with transmission disabled.

  • Prioritize zero impact on the existing bus
  • Fix DE disabled or use a receive-only adapter
  • Do not change termination or biasing
  • Verify any automatic transmit control
  • Treat non-isolated adapters as laboratory tools
Capture optionRole and limit
Commodity tapFast and inexpensive, but weak in timestamp quality, isolation, ring buffering, and drop detection.
RFR Wired TapA later option for enforced read-only capture, hardware timestamps, isolation, ring buffers, and drop detection.

System architecture

Import existing logs, align with human anchors, and convert the failure into a reviewable Clip.

Raw events may receive clip_id, episode_id, and derived_trace_id after reconstruction.

RFR Ingest Agent remains read-only and never blocks the robot control loop.

01

Import

MCAP, CSV, UART dump, application, Linux, modem, and cloud logs.

02

Align

Manual Anchor, lane offset, drift, and explicit timestamp quality.

03

Capture

Range selection, operator marker, or deterministic anomaly creates the Failure Clip.

04

Inspect

Review raw, decoded events, AI summary, evidence jumps, and uncertain order.

Software ingestion stays non-invasive

Local spool, dropped_event_count, and asynchronous transmission are assumed. Import is read-only, and the ROS mirror only subscribes. An active RFR Bridge is a separate component whose safety conditions and BYPASS behavior must be validated per device.

AI Debugger / Read-only analyst

AI never operates the machine.

CRC errors, timeouts, framing errors, and response delays are first detected deterministically.

AI organizes candidate causes, evidence, and recommended checks. It does not replace human review of raw logs.

Raw bytesAuthoritative communication evidence
Frame detectionDeterministic boundaries
Protocol decoderValidated meaning
Rule-based detectionConfirmed anomalies
Structured timelineEvidence and uncertainty
AI summaryGrounded, read-only assistance

Design principles

Build only what gets the team back to the next valuable experiment.

  1. 01Confirm protocol meaning with real logs and real hardware
  2. 02Observe passively before any active connection
  3. 03Require human approval in Shadow Translation
  4. 04Always preserve raw communication as primary evidence
  5. 05Never hide uncertain order or unknown timestamp quality
  6. 06Keep raw logs off BLE Mesh; retrieve selected windows over Direct BLE
  7. 07Keep Wi-Fi and LTE off Nodes; add them only to the Gateway when required
  8. 08Keep AI read-only and grounded in the Evidence Chain