Skip to content

DDS AutoDiag

AI-powered diagnostics for ROS 2: find out why your nodes can't talk, see every broken connection, and let AI guide the fix. Every AI change is checked against your live system first.

Get started See how it works

v0.1.0 ROS 2 Jazzy · Humble Fast DDS · Cyclone DDS · Zenoh-aware Read-only by design Works offline MIT licensed

The problem

Topics show up in ros2 topic list, but no data arrives. RViz stays empty. Two containers see each other's nodes but never exchange a message. The cause is usually one of a dozen settings: a QoS mismatch, a different ROS_DOMAIN_ID, a discovery range, the RMW implementation, or shared memory between containers. Finding it by hand costs hours.

DDS AutoDiag checks all of them in seconds, explains what it found, and tells you exactly what to change.

Step 1 Collect env · RMW · network read-only probes Step 2 Compare two hosts or containers optional peer Step 3 Evaluate rule engine evidence per finding Step 4 Explain next step + docs optional AI Step 5 Verify live sample delivery, not guesses

What you get

  • Rule-based diagnosis

    23 checks for QoS, domains, discovery range, RMW and distro mismatches, DDS XML, shared memory between containers and more. Every finding comes with evidence, a next step and official documentation where one applies.

    Diagnostic rules

  • Live topic inspection

    Publishers, subscribers and their QoS for any topic, incompatibilities highlighted, and a one-shot message sample that proves data actually arrives.

    Web dashboard

  • Configure RViz with AI

    Describe the view you want. The AI plans the displays, AutoDiag validates them against live topics and TF, adds the QoS fixes it diagnosed, and applies only what you confirm. With Undo.

    Configure RViz

  • Grounded AI assistant

    A RAG pipeline over curated official ROS 2, DDS and Docker docs. Answers stream in, cite real sources and remember the conversation. Local (Ollama) or Claude, ChatGPT and Gemini.

    RAG & AI pipeline

  • Compare hosts & containers

    Put two robots, PCs or containers side by side, from redacted snapshot files or a token-protected live collector, and see exactly what differs.

    Compare hosts

  • Safe by design

    Never restarts, reconfigures or publishes anything on its own; only the demo you start publishes, on /autodiag_demo/*. Loopback by default, token login, DNS-rebinding protection, strict CSP, and keys that never reach the browser.

    Security model

What's included
Core Diagnosis (CLI and dashboard) · live topic QoS inspection · RViz2 panel
AI add-ons (optional) Grounded assistant · Configure RViz with AI
Extras Host and container comparison · reports · redacted snapshots · Docker and systemd
Experimental Native C++ endpoint checker · previous UI at /legacy

Experimental features may change or be removed before 1.0.

Three ways to use it

dds-autodiag doctor              # is this machine ready?
dds-autodiag check               # diagnose this host
dds-autodiag inspect /scan       # endpoints, QoS, incompatibilities

Scriptable, with --json output and CI-friendly exit codes. Command line guide

dds-autodiag serve --open        # http://127.0.0.1:8080

Overview, topics, hosts, transport (this host vs the loaded peer), tools (tool check, live rule check, live pair) and the AI assistant in one place. Dashboard guide

cd ros2 && colcon build --packages-select rviz_autodiag_panel
source install/setup.bash && rviz2

Panels → Add New Panel → rviz_autodiag_panel → AutoDiag. Health, topic QoS, the diagnosis graph, AI assistant and Configure RViz, right next to your visualization. RViz2 panel guide

How it fits together

Interfaces Server Engines System CLIdds-autodiag check Web dashboard127.0.0.1:8080 RViz2 panelhealth · topics · AI AutoDiag serverFastAPI · read-onlyREST + streaming Rule engineQoS · domain · discovery RAG + AIlocal or cloud models RViz plannervalidated display plans ROS 2DDSgraph

All interfaces talk to one local server. It probes your system read-only, runs the rule engine, and only reaches out to an AI model when you ask it to. Read the architecture