Autonomy Engine

Phase 5

**Phase 5 — AI Decision Layer (Planned)**

The Autonomy Engine is the core AI decision system that allows robots and agents to operate with minimal human intervention. It evaluates real-world state, applies policies, and issues commands back through the platform.

---

Architecture Overview 

Platform World State
       │
       ▼
┌─────────────────────────────┐
│       Autonomy Engine        │
│  ┌──────────┐ ┌──────────┐  │
│  │  Planner │ │ Executor │  │
│  └────┬─────┘ └────┬─────┘  │
│       │            │        │
│  ┌────▼────────────▼──────┐ │
│  │     Safety Controller  │ │
│  └────────────────────────┘ │
└─────────────────────────────┘
       │
       ▼
Robot Commands → robotics/adapters/*

---

Components

Planner

  • Ingests world state from `platform/world/types.ts → World` model
  • Runs task decomposition against active `Workflow` (see `platform/automation/types.ts`)
  • Emits ordered step sequences for the Executor

Executor

  • Iterates planned steps
  • Dispatches commands to robots via `@mlebotics/robotics-agents → AgentRuntime`
  • Reports step status back as `StepRun` updates

Safety Controller

  • Validates every command before dispatch
  • Enforces configurable policy rules (geofences, velocity limits, battery thresholds)
  • Can halt the entire autonomy loop on policy violation

---

Integration Points

| Platform package | Used by |

|---|---|

| `platform/world` | Planner — reads entity state and world context |

| `platform/automation` | Planner — activates and tracks Workflow runs |

| `robotics/agents` | Executor — issues RobotCommand to AgentRuntime |

| `robotics/adapters/*` | Executor — sends commands to physical hardware |

---

Phase 5 TODOs

<!-- TODO: implement Planner with LLM-based task decomposition -->

<!-- TODO: implement Executor with retry + timeout logic -->

<!-- TODO: implement Safety Controller with configurable policy rules -->

<!-- TODO: expose autonomy session REST/tRPC endpoints in packages/api -->

<!-- TODO: add autonomy status dashboard route in apps/console -->

<!-- TODO: write unit tests for safety policy enforcement -->

---

Safety Design Principles

  • **Fail-safe by default** — any unrecognized state triggers a stop command
  • **Policy-first** — no command is dispatched without safety controller approval
  • **Audit trail** — all decisions are logged as `AuditEvent` (see `platform/shared/types.ts`)
  • **Human-in-the-loop** — Phase 5 ships with an override UI for operators