Curriculum review — 2026-10-13¶
Latest review: the October 14 course-by-course pass covers all 158 current course entry pages, targeted technical corrections, the connected Physical AI Design Studio, and eight original SVG/PNG teaching diagrams. The report below preserves the earlier expansion and its validation history.
This update reviews repository structure across the roadmap and improves the learning and assessment layer at every existing phase/course entry page. The focused technical expansion covers DeepSeek V4.1 Flash NVFP4 inference and NVIDIA Isaac Sim / Isaac Lab.
Scope and evidence¶
The initial inventory covered 534 tracked source Markdown pages, including repository and project documentation. It screened structure, topic coverage, course entry points, and local links. Technical close reading concentrated on the shared authoring standard, phase/track guides, inference course, and Isaac course. This is not a claim that every command in all historical lectures was rerun or that every external resource was revalidated.
The existing uncommitted change in Isaac Lecture 01 was preserved. Generated docs/ and _site/ content is rebuilt from the source pages rather than edited directly.
Findings and changes¶
| Finding | Learning consequence | Change |
|---|---|---|
| Many guides had long concept/resource lists without an explicit assessment | Reading could be mistaken for mastery | Added topic-specific entry diagnostics, baseline labs, variations, required evidence, and exit criteria to 126 course entry pages |
| Existing strong courses used different completion expectations | Learners could not compare progress across tracks | Linked the shared learning workflow from 29 additional course entry pages; 155 existing entries addressed in total |
| Hardware access requirements were mixed with course completion | Learners without a large node could stall or overclaim results | Defined analysis, fixture, simulation, and measured deployment milestones |
| Inference course lacked a V4.1 Flash checkpoint-specific path | Generic MoE/FP4 knowledge did not address actual serialization and compatibility | Added ten lectures and CPU tools for metadata, FP4 representation, and request metrics |
| NVFP4 naming could imply uniform format or automatic memory savings | Incorrect capacity plans or loader assumptions | Distinguished upstream FP4 experts, two community conversions, architecture support, and exact loader support |
| Isaac course needed a progressive validation sequence | Training and deployment failures could hide frame/reset/timing bugs | Added eight lectures, a lab sequence, and executable CPU contract checks |
| Eight-GiB simulation language could read as a supported baseline | Installation and sensor labs could fail before learning began | Clarified supported RTX requirements versus experimental reduced exercises |
| Link checker scanned code as if it were prose and skipped untracked files | False failures and missed new-page errors | Excluded fenced/inline code, added new-file and angle-link coverage, and regression tests |
Coverage by curriculum area¶
| Area | Added or reinforced assessment focus |
|---|---|
| Phase 1 | RTL/testbench behavior, memory/cache reasoning, parallel correctness, synchronization and timing |
| Phase 2 | Board/interface contracts, firmware queues, connectivity recovery, reproducible Linux images, product setup/error behavior |
| Phase 3 | Numerical references, held-out evaluation, calibration and timestamps, model/application failure cases |
| Phase 4A | Constraints, AXI/DMA integrity, timing/CDC, HLS tradeoffs, driver ownership |
| Phase 4B | BSP/deployment identity, sensor/media timing, shared memory, thermal behavior, bounded recovery |
| Phase 4C | Rewrite legality, kernel reference checks, compiler traces, quantization quality, serving latency |
| Phase 5A/B | Topology, collectives, placement, scaling, I/O, library setup versus execution |
| Phase 5C | Edge quality/latency/memory tradeoffs and sustained operation |
| Phase 5D | Frames, assets, partial resets, reward/done semantics, evaluation uncertainty, performance, ROS timing, transfer |
| Phase 5E | Recorded/simulated scenarios, process boundaries, sensor timing, requirements-to-test evidence |
| Phase 5F | Workload-derived budgets, reference models, verification, constrained architecture decisions |
| Phase 5G | Checkpoint identity, numerical correctness, execution traces, per-rank capacity, quality-preserving useful throughput |
| Phase 6 | Evidence-backed explanations, calculations, design changes under a new constraint |
The practice additions use topic-specific assignments and preserve existing lectures. A common format makes expectations visible, but cannot replace a worked explanation of the underlying subject. New inference and Isaac lectures include worked reasoning, failure cases, and self-check guidance to make that teaching concrete.
Technical source boundaries¶
- The official DeepSeek V4.1 Flash model, its configuration, and technical report ground the architecture lessons.
- s-zaizen's conversion is the default community-checkpoint case. Its card does not claim validated serving commands or quality/throughput benchmarks. LibertAI's conversion is a separate format/quality comparison.
- vLLM's V4.1 module establishes architecture implementation evidence; each checkpoint/runtime/backend combination still needs verification.
- Isaac Sim 6.1 requirements and Isaac Lab v3.0.0-EA documentation ground the simulator release family. Learners must save immutable commits and resolved configurations for their experiments.
Reproduce the portable checks¶
From the repository root:
python3 scripts/audit_curriculum.py
python3 scripts/check_links.py
python3 -m unittest discover -s scripts -p 'test_*.py'
python3 -m unittest discover -s 'Phase 5 - Advanced Topics and Specialization/Track G - ML Systems Engineering/AI Inference Engineer 2026/Part 5 - DeepSeek V4.1 Flash NVFP4/labs' -p 'test_*.py'
python3 -m unittest discover -s 'Phase 5 - Advanced Topics and Specialization/Track D - Robotics/Isaac Sim and Isaac Lab/labs' -p 'test_*.py'
STRICT=1 bash scripts/build_docs_local.sh
The audit script emits review signals and can export the complete inventory with --json. Keyword presence is not a quality grade. Link checks validate local file targets; MkDocs also reports anchor/navigation issues. Neither proves external URL availability.
Local validation: 23 portable tests passed, all local file targets resolved, all four CPU demos ran, and the strict MkDocs build passed. Hardware execution was not part of these checks.
The new CPU tests validate metric arithmetic, corrupt/invalid inputs, packing, reset isolation, coordinate checks, timing definitions, uncertainty, and command freshness. They do not establish full-model inference, NVFP4 kernel performance, simulator execution, trained-policy success, or real-robot transfer. Those require the hardware-specific experiments described in the courses.
Learning guide · Authoring standard · Roadmap
Professional VLA expansion — October 13, 2026¶
Added VLA Foundations to Frontier 2026: 32 detailed lectures covering history, grounding, action representations, imitation/diffusion/flow learning, data engineering, adaptation, RL, training infrastructure, evaluation, Isaac integration, asynchronous control, robustness, and current research. Four source dossiers and a dated frontier register distinguish papers from available artifacts and restricted access.
The course adds a 16-week study path, notation/glossary reference, worked examples, negative controls, answered self-checks, and two capstones. Its CPU labs train an inspectable nine-parameter grounding model and check action normalization, episode masks, split groups, scheduling, and paired task statistics. These fixtures do not establish pretrained-model, simulator, or physical-robot performance. New VLA tests are included in curriculum CI; the 23-test result above records the earlier roadmap review, before this expansion.
VLA expansion validation: 35 portable tests pass across the link checker, inference, Isaac, and VLA suites. All four VLA lab-guide shell blocks execute successfully, local Markdown links resolve, and the strict documentation build renders all 32 lectures and their downloadable Python helpers. No GPU, simulator, API, or physical-robot execution was performed.
Programmable VLA accelerator expansion — October 14, 2026¶
Developed the supplied chip proposal into Programmable VLA Inference Accelerator: ten lectures connecting executable workloads, tensor/vector architecture, memory residency, mixed precision, valid reuse, compiler/runtime interfaces, action scheduling, task-level energy, hardware prototyping, and a reviewed SoC proposal.
Two primary-source ledgers record current hardware specifications, model access, GR00T timing boundaries, and 2026 accelerator/quantization research. They preserve unresolved source discrepancies and distinguish publications, available implementations, analytical assumptions, and measurements. The research snapshot ends on October 14; later October developments are outside its scope.
Five CPU labs reproduce the proposal's arithmetic and test assumptions before hardware commitments. Two original, reproducible SVG/PNG figures explain partial residency and scheduling boundaries. The course is linked from the roadmap, VLA capstone, compiler track, physical-AI studio, and chip-design track, with all ten lectures in site navigation.
Portable validation: 51 tests pass across the existing suites and 16 new accelerator tests. The new tests include decimal release/deadline boundaries, partial packed blocks, cold/warm fills, SRAM reserves, tile padding, dispatch cost, failed-attempt energy, and observation age. All local Markdown file targets resolve. These results validate teaching fixtures; no GPU, simulator, FPGA, or silicon measurements were performed for this expansion.
Run the new suite from the repository root: