The plan, version 1.1
The full design.
Everything behind FrankenRobotics in 70 sections and appendices, from the algebra of a single joint to the packet that goes to the workshop.
Where to start
New here
What it is, what it’s built to do, and the scenarios that drive it.
Testing
Adversarial worlds, failure analysis, replay and the test program.
The math
Forty constructions, the algebra core, proofs and worked answers.
Thesis
Part I · 8 entriesThe field as it stands, the gaps in it, and what FrankenRobotics adds.
§24,272 wordsThe ten-fold thesisWhere the time goes in a robot-design loop, and how the lab aims to get ten times faster.
§3838 wordsTwelve principlesThe twelve rules behind every design decision.
§44,860 wordsBets, each with its falsifierThe ideas the project rests on, and the experiment that tests each one.
§5535 wordsReadersWho the plan is written for.
§62,429 wordsGoals, non-goals, anti-goalsWhat FrankenRobotics sets out to do.
§72,493 wordsNorth-Star scenariosThe before-and-after scenarios that define what the lab has to do.
§81,825 wordsThe seven claims and the claim productThe kinds of result the lab reports, and the evidence behind each.
The mathematical spine
Part II · 4 entriesHow algebra and Lie theory shape the whole system.
§1024,772 wordsThe constructions (C-01 onward)Forty mathematical constructions, each with its payoff and its test.
§112,642 wordsConventionsFrames, signs, units and transposes: one convention for each.
§126,313 wordsLemma: proofs, certificates, the offline auditorProofs, certificate bundles, and a small program that re-checks them offline.
The body
Part III · 9 entriesLie groups, spatial algebra, cones and symmetry, in Rust.
§1412,252 wordsSkeleton: design language, modules, archetypes, preflightHow a robot is described: typed modules, eighteen kinds of body and the two-second design check.
§157,354 wordsSinew: actuators, transmissions, powerMotors, gearboxes, batteries and heat, at several levels of detail.
§1616,629 wordsKinesis: dynamics on three tiersRigid-body dynamics at three speeds, from a fast GPU step to a strict referee.
§179,052 wordsPhylum: families of bodies in one batchDifferent bodies stepping together in one GPU batch, with no recompile.
§188,898 wordsTraction: collision, contact, wheels, tracks, terrainCollision, friction, wheels, tracks and soft ground.
§1919,082 wordsHabitat: scenes, living worlds, clearance, retrofitsGenerated and scanned buildings, their hazards, clearance maps, and fixes for the room.
§2012,681 wordsSensorium: sensors, clocks, observabilityCameras, lidar, touch and clocks, and what a robot can tell from them.
§2112,693 wordsSpark: GPU backends and the kernel toolchainKernels for CUDA and Metal, generated from one source and checked against the CPU.
The mind
Part IV · 7 entriesState estimation, model-based control, gaits and safety filters.
§2312,910 wordsCortex: the gym, learning, family policies, inheritanceTraining one policy for a whole family of bodies, on the GPU.
§2415,330 wordsArena: tasks, requirements, fixtures, metrics, qualificationTasks, scoring, and the sealed exam an agent can’t peek at.
§2539,322 wordsAugur: mission queries and inverse designAsk for every robot that can do a mission and get the whole family back.
§266,790 wordsNemesis: adversarial worlds and rare failuresThe search for the world that beats a robot, and the smallest version of it.
§276,879 wordsAutopsy: counterfactual debuggingReplay a fall with one thing changed and find out what caused it.
§2821,486 wordsAtelier: the studio loop, analyses, comparison, lineage, cascade, decision coresThe studio loop: edit, test, compare, bisect, repeat.
The record and the world
Part V · 4 entriesEvery run recorded, replayable and queryable.
§309,527 wordsGround: calibration against realityTuning the model to measurements from real robots.
§3110,601 wordsFoundry: structure, manufacture, lifecycleStructural sizing, tolerances, fatigue life and the build packet.
§3214,628 wordsTether: CLI, daemon, MCP, dashboard, viewerThe command line, the daemon, the MCP server and the viewers.
Substrate and engineering
Part VI · 5 entriesHow the lab builds on FrankenSim, FrankenTorch, FrankenSQLite and the rest of the suite.
§346,855 wordsWorkspace, crates, formatsThe Rust workspace: crates, layers and file formats.
§355,634 wordsNumerics and determinismWhich results are exact, which are repeatable, and to how many digits.
§367,173 wordsThe performance programSpeed targets for every kernel and tier, and how they’re measured.
§373,405 wordsSafety, security, privacySandboxing, permissions and privacy.
Verification, the ten-fold test, gates and risks
Part VII · 5 entriesKnown answers, planted bugs, fuzzing and cross-checks for every subsystem.
§3915,590 wordsThe ten-fold testThe head-to-head experiment that measures the ten-fold goal.
§4026,701 wordsThe gate registerEvery acceptance test in the project, in one list.
§416,128 wordsRisks and the pre-mortemThe risks being watched, and the measurement that gives early warning of each.
§427,283 wordsOpen questionsDecisions still to be made, and the safe default for each.
Delivery
Part VIII · 5 entriesWhat ships when, from the first kernel to the ten-fold test.
§442,870 wordsWork packagesThe deliverables between the milestones and the tasks.
§4517,283 wordsThe beadsThe task list, ready to be worked through.
§461,316 wordsThe first implementation sessionThe first day of building, step by step.
§472,338 wordsDefinition of done and the release rubricWhat has to be true before a release.
Appendices
15 entriesEvery term in the plan, in a line each.
Appendix B11,119 wordsIdea passesWhere the ideas came from and where each one ended up.
Appendix C15,958 wordsEV matrix and the killed-idea registerHow the candidate mechanisms were scored and chosen.
Appendix D6,230 wordsMathematics compendiumThe formulas, all in one place.
Appendix E12,069 wordsWorked cardsThe plan’s main relations with numbers put in, worked by hand.
Appendix F4,101 wordsUpstream work itemsChanges planned for the sibling projects.
Appendix G3,195 wordsReference robots, scene assets, licences and hazard listsThe robots, scenes and licenses used as references.
Appendix H3,245 wordsSource registerSources, with links.
Appendix I5,440 wordsType sketchesThe main Rust types, sketched.
Appendix J5,809 wordsVerb registryEvery command, generated from one registry.
Appendix K1,648 wordsReview log and plan-time recordThe record of reviews and what each one changed.
Appendix L7,940 wordsWorked artifactsNine files from one job: a quadruped carrying a tray through a kitchen.
Appendix M50,586 wordsKnown-answer suiteHundreds of cases with known answers for the test suite to reproduce.
Appendix N30,146 wordsCounterexample catalogueHundreds of tempting shortcuts, each with the small case that breaks it.
Appendix O5,497 wordsRefusal code registryEvery error code, its message and its repairs.
Overview
1 entryDocument control
Part 0 · 8 entriesHow the plan words its requirements and labels its statements.
§0.2320 wordsWhat "implemented" meansThe eight things that have to hold before a feature counts as done.
§0.3381 wordsStable identifier prefixesHow every requirement, gate and construction gets an identifier that never changes.
§0.4320 wordsReading guideWhich sections to read first, depending on what you’re after.
§0.5533 wordsSubsystem namesWhy the subsystems are called Skeleton, Sinew, Cortex and so on.
§0.6765 wordsHow this plan was producedThe four stages of drafting and review, and what each one left behind.
§0.7806 wordsPlan hygiene rulesThe rules that keep a long plan consistent as it changes.
§0.83,180 wordsHow evidence from the planning exercises binds the planWhat was worked out on paper first, and how it shaped the design.