Epic: ADA Feeding Pipeline

[siddhss5]

Goal

Fully functioning feeding demo in MuJoCo that faithfully reproduces ada_feeding's capabilities using plain synchronous Python functions with structured `Outcome` returns. Same code runs on sim (SimContext) and real hardware (HardwareContext).

Architecture

React App (user)  ──┐     ┌── Viser Panel (developer)
                    │     │
              FastAPI/WS  event_loop
                    │     │
         ┌──────────▼─────▼──────────┐
         │   Feeding Logic (pure Py) │
         │   feed_bite(), behaviors  │
         │   No viser, ROS, MuJoCo   │
         └──────────┬────────────────┘
                    │ ExecutionContext
         ┌──────────▼────────────────┐
         │   mj_manipulator          │
         │   servo_to_pose,          │
         │   ft_guarded_move, Outcome│
         └──────┬──────────┬─────────┘
           SimContext   HardwareContext

Implementation Slices

Slice 1: Foundation (mj_manipulator)

• `Outcome` + `FailureKind` structured return types
• `ForceThresholds`, `SpeedProfile` dataclasses
• `servo_to_pose()` — Cartesian servo with deceleration + F/T monitoring
• `ft_guarded_move()` — move until F/T threshold exceeded

Slice 2: Domain + Behaviors (ada_mj)

• `FoodItem`, `AcquisitionSchema`, `ForkState` domain types
• Behaviors: `move_above`, `acquire_food`, `extract_food`, `level_fork`, `transfer_to_mouth`, `retract_from_mouth`
• Each is a plain function returning `Outcome`

Slice 3: Task + Demo (ada_mj)

• `feed_bite()` task — 30-line function composing behaviors
• `feeding_demo()` loop — acquire + deliver each food item
• Food objects + plate in MuJoCo scene

Slice 4: Viser Feeding Panel (ada_mj)

• Developer-facing: food selection, action buttons, F/T plot, status

Slice 5: React App Rewrite (feeding_web_interface)

• User-facing: preserves validated UX for patients with motor disabilities
• Strips ROS/rosbridge, communicates via FastAPI + WebSocket
• Same meal state machine: PreMeal → BiteSelection → RobotMotion → BiteDone

What we preserve from ada_feeding

• F/T thresholds per phase (approach 20N, grasp 15N, extraction 50N, mouth 1N)
• Speed profiles (0.15→0.06 m/s deceleration near mouth)
• Acquisition schemas (tilt angles, twist directions, durations)
• Mouth approach geometry (offset + orientation)
• Safety abort on F/T exceedance
• Accessible UX (large buttons, clear status, tested with real users)

What we eliminate

• py_trees tree construction → plain function calls
• Blackboard data passing → typed function arguments
• `scoped_behavior` / `servo_until` idioms → Python control flow + internal loops
• ROS action servers / rosbridge → direct calls + FastAPI
• MoveIt2 action clients → `arm.plan_to_configuration`
• GNU Screen orchestration → single process