In progress · updated 2026-10-30 · embodied-ai-architect/docs/plan/roadmap-v2.md
Baseline: v0.6.0 (Feb 2026) | Target: First paying customer by Release 1.0
Timeline: 18 months (Mar 2026 – Aug 2027)
Feature Inventory: feature-inventory.md
Prior Roadmap: roadmap.md (original 12-month R&D plan)
Feb 2026 Mar–Apr 2026 May–Jul 2026 Aug–Nov 2026 Dec 2026–Mar 2027
┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────────┐
│ 0.6 │ │ 0.7 │ │ 0.8 │ │ 0.9 │ │ 1.0 │
│ Specs ◄──│──→│ Prod CLI │───→│ Opt Loop │──┐──→│ Co-Sim │──────→│ Customer │
└──────────┘ └──────────┘ └──────────┘ │ └──────────┘ │ Ready │
(parallel if ≥2 eng) Apr–Jun 2027 Jul–Aug 2027
┌──────────┐ ┌──────────┐
│ Enterprise│→ │ Sim2Real │
└──────────┘ └──────────┘
Critical path: 0.7 → 0.8 → 1.0 (optimization loop is the core differentiator)
Parallel path: 0.9 (co-simulation) can proceed alongside 0.8 with a second engineer
Business Value: Structured, versioned requirements capture. Users define what they want to build before the AI architect designs it. Every field change is tracked with provenance (who, when, why), enabling auditable design history from day one.
# Feature Inventory Ref Status Description 1 System Spec model (8 subsystems) 10.1–10.5 DONE Perception, Compute, Power, Sensor, Actuator, Comms, Autonomy, Safety — all Optional for incremental definition 2 Event-sourced store with content-addressed versioning 10.6–10.19 DONE Append-only JSONL log, SHA256 blob storage, thread-safe locking, auto-snapshot after 50 events 3 14 CLI subcommands (branes spec *) 10.22 DONE new, list, show, set, delete, commit, history, diff, tag, export, import, why, validate, resolve 4 6 LLM tools for chat agent 10.23 DONE list/create/read/modify/validate/field_history — agent can reason about and modify specs 5 6 predefined templates 10.20 DONE drone-perception, quadruped-nav, industrial-inspection, amr-warehouse, edge-camera, biped-humanoid 6 7 cross-subsystem validation checks 10.21 DONE Power vs compute, perception latency, platform implications, safety redundancy, mission duration, comms bandwidth, autonomy sensors 7 82 tests (44 model + 38 store) 10.25 DONE Full coverage of models, events, store, diff, templates, validation 8 Design document + tutorial 10.26, 10.27 PARTIAL Design doc complete; tutorial started but needs completion
Feature Inventory Ref Description Multi-objective optimization engine 7.1–7.15 MAP-Elites + Bayesian BO + NSGA-III, MCP server, CLI integration Codebase analysis pipeline 9.24–9.26 Static scanner, LLM 4-pass analyzer, workload profile converter Documentation site 9.5–9.7 Astro Starlight with tutorials, catalog pages, MCP reference CI/CD pipeline 9.1–9.3 Black + Ruff lint, pytest, semantic-release, docs deployment 7 discoverable demos 6.16 SoC designer, DSE Pareto, SoC optimizer, KPU RTL, HITL safety, experience cache, full campaign
Spec → pipeline bridge: SystemSpec doesn’t yet drive the design pipeline (10.28, 10.30)
Spec tutorial incomplete (10.27)
Cross-spec comparison not implemented (10.29)
Timeline: Month 1–2 (Mar–Apr 2026)
Business Value: Tool is usable by early adopters without workarounds. Users can run benchmarks on remote hardware and manage configuration through the CLI instead of editing YAML files.
# Feature Inventory Ref Priority Effort Description 1 Wire SSH/K8s backends into CLI 6.5 P0 M Replace “not yet implemented” fallbacks in workflow.py:101-106 and benchmark.py:88-92 with actual backend dispatch 2 Implement branes config set 6.6, 6.7 P0 S YAML-validated config editing from CLI (replace manual file editing) 3 Implement branes backends add/test/configure 6.8, 6.9, 6.10 P0 M Full backend lifecycle: add SSH host/K8s cluster, test connectivity, configure resources 4 Fix SSH model serialization 4.5 P0 M Replace _save_model_architecture() placeholder with ONNX-based serialization + transfer 5 Fix K8s model reconstruction 4.10 P0 M Replace simplified Sequential placeholder with ONNX model loading from ConfigMap 6 Add Local GPU benchmark backend 4.3 P1 M CUDA benchmarking with memory profiling and GPU utilization 7 Spec → pipeline bridge 10.28, 10.30 P1 M Auto-extract DesignConstraints from SystemSpec; spec drives workload analysis → hardware → PPA 8 Bayesian MOO predict() 7.8 P2 S Implement fitted model persistence and prediction for rapid design exploration 9 Complete spec tutorial 10.27 P2 S Finish the spec-of-specs tutorial walkthrough
branes benchmark run model.pt --backend remote_ssh returns results from a remote machinebranes benchmark run model.pt --backend kubernetes runs on a K8s clusterbranes config set benchmark.timeout 120 updates config with validationbranes backends add my-gpu-box --type ssh --host 192.168.1.100 works end-to-endbranes spec new my-drone --template drone-perception && branes design from-spec my-drone runs full pipeline from specAll 640+ existing tests still pass
Risk Probability Impact Mitigation ONNX serialization doesn’t cover all model architectures Medium High Support PyTorch scripting as fallback; document supported model types K8s ConfigMap size limit (1MB) for large models Medium Medium Use PersistentVolumeClaim for models >1MB
Timeline: Month 3–5 (May–Jul 2026)
Business Value: The agent autonomously improves designs — the core differentiator that separates this from static analysis tools. Customers see the AI architect iterate and converge on better designs.
# Feature Inventory Ref Priority Effort Description 1 Iterative optimization loop 1.30 P0 XL Main loop: optimizer → ppa_assessor → critic → re-plan. Convergence detection, max iterations, improvement threshold 2 Experience cache (SQLite + similarity) 1.31, 1.32 P0 L Persistent storage of past designs with nearest-neighbor lookup for warm-starting new explorations 3 Human-in-the-loop approval blocking 9.18 P0 M Replace auto-approve placeholder with real blocking gate (CLI prompt + API endpoint for future web UI) 4 Dispatcher retry policies 1.39 P1 M Transient failure recovery with exponential backoff; configurable retry count per task type 5 Dispatcher streaming output 1.40 P1 M Stream task progress to CLI/API consumers; real-time optimization status 6 Expanded critic vocabulary 1.21 P1 M Add thermal analysis, reliability (MTTF/electromigration), supply chain risk critique 7 Surrogate model training 7.16 P1 L Train lightweight surrogate from experience cache for 100x faster design space exploration 8 DAG dataflow topology 9.23 P2 M Extend bandwidth analysis from linear chain to arbitrary DAG (parallel branches, fan-in/fan-out)
Agent takes a use-case description, iterates through ≥3 design alternatives, converges on a design meeting all constraints
≥20% improvement over naive single-pass baseline on a standard benchmark goal
Experience cache persists across sessions; warm-started exploration converges 2x faster
Human approval gate blocks on safety-critical actions; CLI shows approval prompt
Risk Probability Impact Mitigation Optimization loop diverges (oscillates between strategies) Medium High Monotonicity check: reject iterations that regress on primary metric Experience cache grows unbounded Low Medium TTL + LRU eviction; archive old designs Surrogate model inaccuracy misleads search Medium Medium Validation against full PPA assessor every N iterations Ships as two disjoint loops (population MOO + single-design dispatcher) with two state schemas; specialists stay deterministic, critic feedback stays coarse High High Follow-on Loop Convergence milestone (roadmap-loop-convergence.md) unifies on one DesignState, agent-ifies critic/optimizer, wires architect skills to a real code iteration
Requires Release 0.7 (CLI must work for testing the loop end-to-end)
Followed by the Loop Convergence milestone (roadmap-loop-convergence.md) — slots between 0.8 and 0.9; unifies the two loops into one multi-agent loop before co-simulation consumes the design state
Timeline: Month 6–9 (Aug–Nov 2026)
Business Value: Designs are validated against physics simulation. Customers can trust that the predicted power/latency numbers match reality within ±15%.
# Feature Inventory Ref Priority Effort Description 1 AirSim/Gazebo ↔ compute simulator bridge 2.1, 2.2, 2.3 P0 XL Bi-directional bridge: physics simulator sends sensor data, compute simulator returns actuator commands, power/latency instrumented 2 Per-component energy instrumentation 2.4 P0 L Attribute energy to perception, planning, control subsystems during simulation 3 End-to-end latency correlation 2.5, 2.6 P0 L Timestamp synchronization between physics and compute; sensor-to-actuator latency measurement 4 Hardware model calibration (Jetson Orin) 2.7, 2.8 P0 L Run physical benchmarks on Jetson Orin; calibrate PPA models to ±15% accuracy 5 Mission profile library — P1 M Standard scenarios: hover, cruise, obstacle avoidance, search pattern, emergency landing 6 Software configuration matrix (50+ configs) 2.10 P1 L Combinatorial sweep: model variants × quantization × scheduling × rate parameters 7 Parallel simulation infrastructure 2.9 P1 L Run 10+ simulations concurrently for dataset generation (10K flight minutes target) 8 External power monitoring (SCPI/USB) 4.14 P2 M Integration with Keithley/NI USB power meters for calibration measurements
Run a simulated drone mission (5 minutes, obstacle avoidance scenario)
Get per-component energy breakdown (perception X%, planning Y%, control Z%)
Predicted vs measured energy on physical Jetson Orin is within ±15%
Sensor-to-actuator latency measured and correlated with compute load
Risk Probability Impact Mitigation AirSim deprecation/maintenance issues High High Support Gazebo as primary; AirSim as secondary. Abstract behind simulator interface Simulation accuracy < ±15% target Medium High Iterative calibration with increasing fidelity; publish confidence intervals Parallel sim requires expensive GPU fleet Medium Medium Start with CPU-only physics; add GPU acceleration incrementally
Requires Release 0.7 (remote backends for running sims on cluster)
Can run in parallel with Release 0.8 (independent engineering track)
Benefits from the Loop Convergence milestone (roadmap-loop-convergence.md): co-sim validates against a single unified DesignState, so co-sim instrumentation is simpler if Loop Convergence lands first (not a hard blocker)
Timeline: Month 10–12 (Dec 2026–Mar 2027)
Business Value: First paying customer. Production-quality tool with professional output, documentation, and CI/CD gating. A customer can install, analyze their workload, get recommendations, benchmark, validate, and receive reports.
# Feature Inventory Ref Priority Effort Description 1 PDF report export 1.26 P0 M WeasyPrint/ReportLab HTML→PDF with executive summary, charts, appendix 2 Comparison reports (design A vs B) 1.27 P0 M Side-by-side PPA comparison with delta annotations and recommendation 3 API reference auto-generation 9.8 P0 M Sphinx/MkDocs autodoc from docstrings → docs site integration 4 Integration tests in CI 9.4 P0 L Separate CI stage for [remote], [kubernetes], [chat] optional dep groups 5 Test coverage reporting (80%+ gate) 9.10 P1 M pytest-cov with coverage badge and CI gate 6 RISC-V deployment target 5.6 P1 L RISC-V cross-compilation + runtime integration 7 FPGA deployment target (Vivado/Quartus) 5.7 P1 XL Bitstream generation from RTL templates; placement + routing 8 Multi-accelerator heterogeneous mapping 1.10 P1 L Architecture composer splits workload across CPU + GPU + KPU simultaneously 9 Production multi-rate framework 8.24–8.26 P1 L Graduate prototype to core package; real-time scheduling with priority inheritance 10 LiDAR drivers (Velodyne, Ouster) 8.13, 8.14 P2 M Production-quality VLP-16/32 and Ouster OS1/OS2 integration
Customer can pip install branes[all] and run complete workflow
branes codebase assess /path/to/project --hardware jetson_orin --power-budget 15 produces PDF reportAll public APIs have generated reference documentation
CI runs integration tests for all optional dependency groups
Test coverage ≥80%
Risk Probability Impact Mitigation FPGA synthesis is unbounded complexity High Medium Start with Lattice iCE40 (small, open-source toolchain); defer Xilinx/Intel to 1.1 Coverage gate blocks releases Low Medium Gradual ramp: 60% at 1.0-rc1 → 80% at 1.0 GA
Requires Release 0.8 (optimization loop is the core product)
Benefits from Release 0.9 (co-simulation results in reports)
Timeline: Month 13–15 (Apr–Jun 2027)
Business Value: Enterprise readiness — team collaboration, regulated-industry compliance, commercial EDA integration. Unlocks defense/aerospace/medical device customers.
# Feature Inventory Ref Priority Effort Description 1 Team collaboration — P0 XL Shared design specs, reports, and optimization results across team members (Git-backed or server-backed) 2 Commercial EDA wrappers 3.30 P0 L Synopsys Design Compiler and Cadence Genus integration for production ASIC flows 3 IP block licensing model + catalog 3.34 P1 L Standardized IP catalog with licensing terms, integration complexity, and qualification status 4 Safety certification compliance — P1 XL IEC 62304 (medical), ISO 26262 (automotive), DO-178C (aerospace) artifact generation 5 Stereo vision fusion (bilateral/guided filters) 8.20 P2 M Replace basic averaging with production-quality depth filtering 6 HDF5 recording/replay 8.29 P2 M Drone perception Phase 4: synchronized multi-sensor HDF5 recording with indexed replay 7 Volume discounts + packaging cost 3.24, 3.25 P2 S Volume pricing curves, chiplet/2.5D/3D packaging cost models
Team of 3 can share design workspace with conflict resolution
RTL synthesis runs through Synopsys DC with timing reports
Generated documentation artifacts satisfy IEC 62304 Class B traceability requirements
Enterprise customer can use the tool in a regulated environment with audit trails
Requires Release 1.0 (customer-ready baseline)
Timeline: Month 16–18 (Jul–Aug 2027)
Business Value: Proof that the co-design thesis works. Measurable, published improvement on real hardware. This is the demo that closes enterprise deals.
# Feature Inventory Ref Priority Effort Description 1 Sim-to-real accuracy (±10% predicted vs measured) 2.12 P0 XL Physical drone with instrumented power/latency; compare against simulation predictions 2 Physical drone deployment + flight test — P0 XL End-to-end: co-designed system running on physical drone in controlled environment 3 ≥50% flight time improvement proof point — P0 L A/B test: baseline drone vs co-designed drone, same mission profile 4 Blueprint package (binary + HW config + BoM + report) — P0 M Deliverable package that a customer can hand to a contract manufacturer 5 NoC topology exploration 3.12 P1 L Move beyond hardcoded 2D mesh; explore ring, tree, and custom topologies 6 RTL timing closure flow 3.18 P1 XL Static timing analysis integration for sign-off quality RTL 7 LiDAR drivers (Livox, ROS bridge, file formats) 8.15, 8.16 P2 M Complete the sensor driver matrix 8 Learned workload estimation 1.2 P2 L Replace keyword-based heuristics with ML model trained on workload corpus
Demonstrate a drone that flies ≥50% longer on co-designed HW/SW vs baseline
Predictions match reality within ±10% on energy and ±15% on latency
Blueprint package accepted by contract manufacturer for quoting
Published case study with quantified results
Requires Release 0.9 (co-simulation must be calibrated)
Requires Release 1.0 (reporting and documentation for publication)
Size Description Approximate Duration (1 engineer) S Small — single file, well-understood change 1–3 days M Medium — 2-5 files, some design decisions 1–2 weeks L Large — cross-cutting, new subsystem integration 2–4 weeks XL Extra large — new domain, research-grade, multiple subsystems 4–8 weeks
┌────────────────────────────────────────┐
│ embodied-schemas (shared catalog) │
│ graphs (roofline, calibration) │
└───────────┬────────────────────────────┘
┌───────────▼───────────┐
│ • Model serialization │
└───┬───────────┬───────┘
┌────────────▼──┐ ┌───▼──────────────┐
│ Release 0.8 │ │ Release 0.9 │
│ Opt Loop │ │ Co-Simulation │
│ • Iterative │ │ • AirSim/Gazebo │
│ • Experience │ │ • Calibration │
│ • Governance │ │ • Energy tracing │
└────────┬──────┘ └───┬──────────────┘
┌────────▼───────────┐ ┆ (benefits: unified
│ Loop Convergence │┄┄┘ DesignState)
The longest dependency chain determines the minimum timeline:
0.7 → 0.8 → Loop Convergence → 1.0 (~12 months) — The optimization loop is the core differentiator, and Loop Convergence (roadmap-loop-convergence.md) is what turns it from LLM-supervised search into a unified multi-agent design loop. It sits on the critical path to a credible customer release.0.9 runs in parallel with 0.8 + Loop Convergence if there’s a second engineer on co-simulation; it consumes the unified DesignState so it benefits from Loop Convergence landing first (soft, not hard, dependency).1.1 → 1.2 is sequential — enterprise features enable the customer relationships needed for sim-to-real hardware access.
Engineers Timeline to 1.0 Notes 1 16 months Sequential: 0.7 → 0.8 → Loop Convergence → 0.9 → 1.0 2 13 months Parallel: (0.8 → Loop Convergence) ‖ 0.9, then converge at 1.0 3 11 months Also parallelize RISC-V/FPGA targets
Risk Releases Affected Probability Impact Mitigation Optimization loop divergence 0.8 Medium High Monotonicity check, max iterations, human escalation AirSim project deprecation 0.9, 1.2 High High Gazebo as primary simulator; abstract behind interface Sim-to-real gap exceeds ±15% 0.9, 1.2 Medium High Iterative calibration; domain randomization; publish confidence intervals FPGA synthesis complexity explosion 1.0 High Medium Start with open-source iCE40; defer Xilinx/Intel Commercial EDA licensing cost 1.1 Medium Medium University program licenses; mock backends for CI Physical drone test failures 1.2 Medium High Incremental: bench test → indoor flight → outdoor flight Single-engineer bottleneck All High High Prioritize 0.7 → 0.8 → 1.0 path; defer 0.9 if needed
This roadmap supersedes roadmap.md (the original 12-month R&D plan). Key changes:
Original Plan This Roadmap Rationale Phase 1: Co-Simulation (Month 1-4) Release 0.9 (Month 6-9) Deferred — CLI and optimization loop have higher customer value Phase 2: Software Architect (Month 5-8) Release 0.8 (Month 3-5) Pulled forward — this is the core differentiator Phase 3: Hardware Co-Design (Month 9-12) Release 1.2 (Month 16-18) Pushed back — need co-simulation + customer base first No CLI/DX milestone Release 0.7 (Month 1-2) Added — current CLI has stubs that block early adoption No customer-ready milestone Release 1.0 (Month 10-12) Added — explicit first-customer gate No enterprise milestone Release 1.1 (Month 13-15) Added — safety certification unlocks regulated industries
All original Phase 1–3 acceptance criteria are preserved in the feature inventory and mapped to releases:
Phase 1 acceptance criteria → Release 0.9 exit criteria
Phase 2 acceptance criteria → Release 0.8 exit criteria
Phase 3 acceptance criteria → Release 1.2 exit criteria
Quarter Release Business Milestone Q1 2026 0.6 Structured requirements capture: specs with versioning, validation, and provenance Q2 2026 0.7 Early adopter program: 3-5 users running benchmarks on remote hardware Q3 2026 0.8 Technical differentiator demo: AI architect iterates and improves designs autonomously Q4 2026 0.9 Validation credibility: show ±15% prediction accuracy against physical hardware Q1 2027 1.0 First paying customer. Professional reports, documentation, CI/CD quality.Q2 2027 1.1 Enterprise pipeline: regulated industry customers (defense, medical, automotive) Q3 2027 1.2 Published proof point: 50%+ flight time improvement, ±10% prediction accuracy