Overview¶

MOSAIC’s rendering subsystem delivers environment frames to the screen through two independent data paths: a Fast Lane for real-time display and a Slow Lane for durable telemetry, plus a pluggable Rendering Strategies layer that adapts pixel, grid, and board-game payloads to Qt widgets. All rendering surfaces converge inside Render Tabs (the central QTabWidget) through individual Training Views (FastLaneTab, LiveTelemetryTab, OperatorRenderContainer, MultiOperatorRenderView).

        %%{init: {"flowchart": {"curve": "linear"}} }%%
graph TB
    subgraph Sources["Frame Sources"]
        W1["CleanRL Worker"]
        W2["XuanCe Worker"]
        W3["Ray RLlib Worker"]
        W4["BALROG Worker"]
        W5["Session Controller<br/>(human / eval)"]
    end

    subgraph FastLane["Fast Lane · shared memory"]
        FLW["FastLaneWriter<br/>SPSC Ring Buffer"]
        FLC["FastLaneConsumer<br/>~60 Hz poll · 16 ms timer"]
        FLT["FastLaneTab<br/>Qt Quick / QML"]
    end

    subgraph SlowLane["Slow Lane · gRPC → SQLite"]
        PROXY["TrainerTelemetryProxy<br/>JSONL → gRPC"]
        GRPC["TrainerService (gRPC)"]
        BUS["RunBus<br/>pub-sub · 2 048 queue"]
        LTC["LiveTelemetryController<br/>background thread"]
        CM["CreditManager<br/>200 credits/stream"]
        RSR["RenderingSpeedRegulator<br/>100 ms · ~10 FPS"]
        LTT["LiveTelemetryTab"]
    end

    subgraph Strategies["Rendering Strategies"]
        REG["RendererRegistry"]
        RGB["RgbRendererStrategy<br/>mouse capture · scroll"]
        GRID["GridRendererStrategy<br/>FrozenLake · Taxi · Cliff"]
        BG["BoardGameRendererStrategy<br/>Chess · Go · Connect4 · …"]
    end

    subgraph Views["Render Views"]
        RT["RenderTabs<br/>(QTabWidget)"]
        ORC["OperatorRenderContainer"]
        MORV["MultiOperatorRenderView<br/>dynamic grid layout"]
    end

    W1 & W2 & W3 --> FLW --> FLC --> FLT
    W1 & W2 & W3 & W4 --> PROXY --> GRPC --> BUS
    BUS --> LTC
    LTC --> CM --> RSR --> LTT
    W5 --> LTT

    LTT --> REG
    REG --> RGB & GRID & BG

    FLT --> RT
    LTT --> RT
    ORC --> MORV --> RT

    style FastLane fill:#e8f5e9,stroke:#2e8b57,color:#333
    style SlowLane fill:#e3f2fd,stroke:#1565c0,color:#333
    style Strategies fill:#fff3e0,stroke:#e65100,color:#333
    style Views fill:#f3e5f5,stroke:#7b1fa2,color:#333

Why Two Lanes?¶

Lane	Mechanism	Use Case
Fast	SPSC shared-memory ring buffer (magic `FLAN`, sequence-number consistency). `FastLaneConsumer` polls every 16 ms (~60 Hz). The first shared-memory live visualization path in RL.	Live training frames: zero serialisation, lossy by design. CleanRL, XuanCe, Ray RLlib, SB3, SBX, Tianshou, and TorchRL workers produce fast-lane frames.
Slow	gRPC → `RunBus` (queue 2 048) → `LiveTelemetryController` → `CreditManager` (200 credits/stream) → `RenderingSpeedRegulator` (100 ms drain).	Durable step/episode telemetry, replay, W&B / TensorBoard integration. All worker types, including BALROG, publish here.

Both lanes converge in Render Tabs: FastLaneTab for the fast lane, LiveTelemetryTab for the slow lane, and OperatorRenderContainer for multi-operator evaluation. The Rendering Strategies layer (RendererRegistry) decides how each frame is painted: pixel, grid, or board game.

Relationship to Other Sections¶

Overview: system-level layer diagram that positions rendering within the broader MOSAIC stack.
Overview: the Log Constants file defines LOG_* codes emitted by rendering components (LOG1001–LOG1015 for operator UI, LOG700+ for workers).
Application Constants: RenderDefaults (min_delay_ms=10, default_delay_ms=100, queue_size=32), BufferDefaults, and telemetry queue sizes all live in gym_gui/constants/constants_ui.py and constants_telemetry.py.