In progress · updated 2026-10-30 · embodied-ai-architect/docs/plan/roadmap-summary.md
Timeline: Months 1-4
Goal: Establish trusted simulation environment where software, hardware, and physics meet
| A/C | Criterion | Metric |
|---|
| 1.1.1 | Simulator Integration | AirSim/Gazebo ↔ CSim bi-directional flow |
| 1.1.2 | Energy Traceability | Per-component breakdown (perception, planning, control) |
| 1.1.3 | Latency Correlation | End-to-end timing sensor → actuator |
| 1.1.4 | Hardware Model Accuracy | ±15% vs physical Jetson Orin |
| A/C | Criterion | Metric |
|---|
| 1.2.1 | Dataset Scale | 10,000+ simulated flight minutes |
| 1.2.2 | Configuration Coverage | 50+ software configurations |
| 1.2.3 | Label Quality | Automated annotation (outcome, energy, latency) |
| 1.2.4 | Correlation Discovery | 3+ actionable insights |
Monthly Breakdown:
| Month | Deliverable |
|---|
| 1 | Simulator integration working |
| 2 | Energy attribution per component |
| 3 | Calibration to ±15% accuracy |
| 4 | 10,000 flight minutes dataset |
Timeline: Months 5-8
Goal: Agent actively optimizes software pipeline on fixed hardware
| A/C | Criterion | Metric |
|---|
| 2.1.1 | Constraint Enforcement | Zero critical constraint violations |
| 2.1.2 | Energy Maximization | ≥20% improvement in Perf/Watt |
| 2.1.3 | Architecture Autonomy | 3+ distinct architectural changes per run |
| 2.1.4 | Failure Interpretation | Textual root cause + proposed correction |
Example Translations:
| User Input | Agent Action |
|---|
| “Maximize flight time, BoM < $350” | Filter HW by cost, prioritize INT8, smaller cache |
| “Detect objects at 10m in any weather” | Prohibit aggressive quantization, add domain randomization |
| “Control loop ≥ 250 Hz” | Enforce <4ms latency, dedicate CPU core |
| “Operate at 60°C passive cooling” | Cap TDP at 5W, prioritize NPU over GPU |
Timeline: Months 9-12
Goal: Agent designs custom hardware to fit optimized software
| A/C | Criterion | Metric |
|---|
| 3.1.1 | Architectural Generation | Modify ≥2 HW params (cache, NPU array) |
| 3.1.2 | Co-Design Efficiency Uplift | +30% Perf/Watt vs Phase 2 baseline |
| 3.1.3 | Constraint-Aware Design | Respect power budget (e.g., 8W TDP) |
| 3.1.4 | Bottleneck Resolution | 40% reduction in primary bottleneck |
| A/C | Criterion | Metric |
|---|
| 3.2.1 | Deployment Success | Boot and complete mission on physical drone |
| 3.2.2 | Sim-to-Real Accuracy | Within ±10% of predicted energy |
| 3.2.3 | Efficiency Proof | ≥50% flight time increase vs baseline |
| 3.2.4 | Blueprint Package | Binary + HW config + BoM + validation report |
| Quarter | Focus | Success Metric |
|---|
| Q1 | Build “Simulated Lab” | Energy model accuracy |
| Q2 | Agent optimizes Software | Energy saved on fixed HW |
| Q3 | Agent optimizes Hardware | Flight time improvement |
| Q4 | Functional validation | SoC functional correctness |
| Role | Phase | Core Deliverable |
|---|
| Lead Systems Architect | 1-3 | Constraint design, blueprint sign-off |
| Agentic AI/RL Engineer | 1-3 | Orchestrator Agent, reward functions |
| Simulation & Tooling Engineer | 1-3 | Co-Sim pipeline, log interpretation |
| Embedded/FPGA Specialist | 1-3 | Optimized kernels, physical deployment |
| Custom HW Synthesis Engineer | 3 | RTL/Verilog from Agent output |
| V&V Engineer | 3 | Sim-to-Real accuracy validation |
Budget: $1.5M - $2.0M (12 months)
| Phase | Timeline | Deliverables |
|---|
| 1. Foundation | Weeks 1-2 | Astro Starlight site, landing page, getting started |
| 2. Catalog | Weeks 3-4 | Auto-generated pages from embodied-schemas |
| 3. Tutorials | Weeks 5-8 | 7 step-by-step guides |
| 4. Reference | Weeks 9-10 | CLI, MCP tools, API docs |
| 5. Interactive | Weeks 11-12 | Hardware finder, compatibility matrix |
Tech Stack: Astro Starlight + llms.txt for AI discovery
| Phase | Timeline | Deliverables |
|---|
| 1. Foundation | Weeks 1-2 | Convert YAML → markdown chunks, vector index |
| 2. Analytical | Weeks 3-4 | graphs roofline integration, text-formatted tools |
| 3. Knowledge | Weeks 5-8 | 15 deployment guides, 10 optimization recipes |
| 4. Benchmarks | Weeks 9-10 | 30 measured model-hardware benchmark entries |
Architecture: LLM-native (text in context, not SQL)
| Risk | Probability | Severity | Mitigation |
|---|
| Simulation Accuracy | High | High | Empirical calibration, Design Rule Checks |
| Agent Hallucination | High | High | Hard-coded DRC rejects invalid outputs |
| Long Iteration Cycles | Medium | High | Surrogate models for fast search |
| Sim-to-Real Gap | Medium | High | Domain randomization (wind, noise, drift) |
| Vendor Lock-in | Medium | Medium | HAL for ARM, RISC-V, FPGA backends |
The Embodied AI Architect delivers:
- Autonomous Design: Agent generates optimized HW/SW configurations from natural language constraints
- Validated Efficiency: 50%+ flight time improvement, verified on physical hardware
- Blueprint Package: Deployable binary + HW config + BoM + validation report
- Documentation: Complete user docs, tutorials, and LLM-integrated knowledge base