Domain Profiles

ROBIN is domain-agnostic. The same stack works for welding, spray coating, machining, or any robotic manufacturing process. Switching domains is a single environment variable - no code changes, no rebuild.

Switching Profiles

ROBIN_PROFILE=spray_coating docker compose up -d

The Alert Engine loads the new profile at startup. The dashboard fetches it at runtime via GET /profile, so labels, units, and AI config update automatically.

Default profile: welding.

Profile Comparison

Concept	Welding (default)	Spray Coating
Process name	Weld	Coat Job
Height metric	Bead Height (mm)	Coating Thickness (mm)
Width metric	Bead Width (mm)	Coverage Width (mm)
Primary AI inputs	Wire Feed Speed (m/min), Travel Speed (m/s), Arc Length Correction (mm)	Line Speed (mm/s), Flow Rate (ml/min), Nozzle Pressure (bar)
Auxiliary measured telemetry	Wire Speed, Welding Current, Arc Voltage	Line Speed, Flow Rate, Nozzle Pressure
AI default mode	parameter_driven	geometry_driven

Canonical Profile Demos

Run the robust dual-mode demos directly from profile scripts. Each script will create the process, configure AI expectations, then wait for you to press Start from the dashboard before streaming data (pass --no-prompt to skip the wait):

# Welding (default profile)
python demo/profiles/welding_profile.py --mode both --duration 120 --interval 2
# -> select the process in the dashboard and press Start

# Spray coating profile
ROBIN_PROFILE=spray_coating docker compose up -d
python demo/profiles/spray_coating_profile.py --mode both --duration 120 --interval 2
# -> select the process in the dashboard and press Start

How Profiles Work

        graph TD
    subgraph profiles ["config/profiles/"]
        W["welding.yaml"]
        S["spray_coating.yaml"]
    end
    ENV["ROBIN_PROFILE\nenv var"] -->|selects| LOADER["ProfileLoader"]
    W --> LOADER
    S --> LOADER
    LOADER -->|"vocabulary\n+ skills"| ENDPOINT["GET /profile"]
    LOADER -->|"model_path"| AILOAD["Load AI\ncheckpoint"]
    LOADER -->|"feature_order\n+ tolerance"| DEV["Deviation\ndetection"]
    ENDPOINT -->|"runtime fetch"| DASH["Dashboard renders\ndomain labels"]
    

Profile files live in config/profiles/ and are plain YAML. A profile contains:

• vocabulary - dashboard labels for the profile’s process and telemetry terms

• fields - telemetry field display names and units

• ros2.topics - ROS 2 topic names for geometry, process parameters, pose

• skills - robot capabilities (service/action names, types, descriptions)

• ai - feature order, model path, default tolerance, default mode, inverse bounds, and inverse-optimizer settings

• dds - DDS bridge topic mapping

See config/profiles/welding.yaml for a complete example.

Per-Profile AI Models

Each profile specifies its model checkpoint in the ai.model_path field. The reference welding profile points at a committed, trained checkpoint, so its AI path works out of the box. The spray-coating profile points at data/models/spray_coating/process_geometry_mlp.pt, which is not committed - so the spray demo runs the configuration-reuse and deviation-against-target path rather than an AI-prediction path.

Note

If a profile’s ai.model_path is missing on disk, the engine falls back to whatever *.pt it can find, which may be another profile’s model with different feature names. Train and commit a per-profile model to enable the AI path honestly:

poetry run python scripts/train_profile_model.py --profile spray_coating
# writes the configured ai.model_path when it is missing

The same reuse procedure applies when running without a model: the deviation reference falls back to the target geometry instead of an AI prediction (as in the spray-coating profile below).

Creating a New Profile

In short: copy config/profiles/welding.yaml, edit the vocabulary, fields, ros2.topics, skills, ai, and dds sections for your domain, optionally train a per-profile model, and launch with ROBIN_PROFILE=<name> docker compose up -d.

This page is the canonical reuse procedure - the configurable-vs-welding-specific boundary, the welding -> spray-coating worked example, telemetry/topic/skill mapping, and the no-model path are all covered above. See also demo/profiles/README.md for a side-by-side profile comparison.