Motion Autocomplete Architecture

Motion Autocomplete Architecture

## Objective
Predict a near-future physical intent (sub-second horizon) and execute environment preparation actions before movement completion.

## Runtime Pipeline
1. Ingestion (`src/ingestion`)
- Simulated multimodal stream generator emitting a `MotionFrame` per timestep.
- Modalities: IMU acceleration/gyroscope, pose reach vector/torso tilt/wrist height, local environment metadata.
2. Fusion Prediction (`src/engine/predictor.py`)
- Sliding window over recent frames (`window_size`, default 8).
- Computes aggregated features: vertical acceleration, turn rate, forward reach, wrist height, torso tilt, planar acceleration.
- Scores intents and selects max-confidence class:
- `lifting_object`
- `turning_around`
- `reaching_for_object`
- `walking_towards_exit`
- `idle`
3. Context Preparation (`src/context/orchestrator.py`)
- Applies a minimum confidence gate.
- Enforces per-intent cooldown to prevent repeated rapid triggers.
- Emits structured actions (`target`, `operation`, `priority`, `reason`) for downstream systems.
4. Presentation (`src/main.py`)
- Human-readable console trace or JSON events for integration.

## Data Contracts
- `MotionFrame`: combines `MotionData`, `PoseData`, and `EnvironmentContext`.
- `PredictedIntent`: includes `intent_label`, `confidence`, `horizon_ms`, and feature metadata.
- `OrchestrationResult`: includes status (`prepared` / `suppressed_*`) and context actions.

## Configuration
`config/default.json`
- Predictor: `window_size`, `prediction_horizon_ms`
- Orchestrator: `min_confidence`, `cooldown_seconds`

## Extension Points
- Replace heuristic fusion scores with trained sequence model (LSTM/Transformer).
- Feed live streams from Graph Kernel/Perception Mesh endpoints.
- Persist predictions/actions to Supabase for online adaptation.