# Decisions

Dated log of Steve's Hux decisions as they landed on `main`. Newest intent **supersedes** older language on the same topic.

**Standing rule (Steve):** merge docs PRs promptly; keep `main` current; **log each merge here.** Do not leave conflicting docs PRs open against stale `main`.

Source of truth is this repo (`nova-centauri/hux-robot`). Docs only. **No spend.** No SKU locks beyond what is already decided below.

## Current governing decisions — V1-PROOF, 2026-09-28 revision B

The user extended V1-PROOF to dependable forward/reverse traversal, left/right turns, turning in place, bounded pushes/uneven terrain and concrete component decisions. This extends the proof scope while retaining its cost, mass and four-axis limits. Earlier dated entries remain history.

| Topic | Current decision |
| --- | --- |
| Active model | V1-PROOF; STAIR-V1 remains parked |
| Cost / mass | Strictly < $1,000; **$900** allocation including $225 reserves; 2.5 kg target / 3.0 kg maximum |
| Actuators | **2 × Pololu 4752 wheels + 2 × ST3215 12 V variant legs through 3:1 belts**; first mobility with pinned 30° legs |
| Drivers | **2 × Pololu 4035 DRV8874**, configured current limiting, sleep/kill/watchdog, PH/EN and fixed-off-time regulation |
| Control / sensing | **Pico 2 + LSM6DSOX SPI**, encoder/current/fault/battery feedback; confirmed equivalent reuse may substitute after qualification |
| Power | 3S approximately 2.2 Ah; **9 V servo / 5 V logic branches**; actual pack, regulators and protection require measured circuit qualification |
| Cameras | **Zero onboard in V1**; existing external video for trials; perception is a later version gate |
| Motion | 0.25 m/s cruise, 0.15 m/s uneven fixtures, 0.4 rad/s arc / 0.6 rad/s pivot; 0.5 m/s not yet qualified |
| Disturbances | Qualification targets: 0.8 N·s longitudinal / 0.4 N·s lateral at 0.20 m, 50/200 ms pulses; 3° grades/cross-slopes, 5 mm smooth bump, 3 mm seam |
| Evidence | Pinned-leg 3D closed-loop simulation with seed/parameter sweep and failure cases; physical validation and powered-leg dynamics remain open |
| Next versions | Carry measured actuator/sensor models, telemetry, fault behavior and tests forward; no promised stair upgrade of this chassis |
| Orders / publication | Model, simulation and component decisions authorized. No purchases or deployment made by this revision |

See [hardware decisions](v1-proof-hardware.md), [simulation evidence](v1-proof-simulation.md), [physical acceptance](v1-proof-validation.md), [PF requirements](requirements.md) and [numerical source](../tools/v1-proof/model.json). Hardware selections are the current engineering baseline; they remain subject to explicit bench rejection criteria and budget limits.

---

## Log

### 2026-09-20 — #1 scaffold (merged)

Initialize Hux wheeled-biped docs scaffold and requirements. SoT is this GitHub repo. North star: climb / descend **~9.5"** residential stairs one wheeled leg at a time. Empty `cad/`, `firmware/`, `software/`. MIT license. No spend. FC not locked.

### 2026-09-20 — #2 XRobots research (merged)

Research-first study packet: [XRobots shortlist](research/xrobots.md), [study plan](research/study-plan.md) Phases A–D. Watch / extract before any print. Do not vendor XRobots trees. RobotX is GPL3 — do not relicense Hux (MIT) without Steve. Printable wheel-leg is **Phase D**.

### 2026-09-20 — #3 Roadrunner + FrRonconi inspiration (merged)

Steve's X-share inspirations: [RAI Roadrunner](research/roadrunner.md) (lab RL wheeled biped — capability existence, not a Hux stack) and the [FrRonconi student two-leg/wheel balancer](research/inspiration.md) (maker-scale vibe). Watch; do not build from the clips.

### 2026-09-20 — #4 electric / 4S / wheel-actuator intent (merged)

**Electric-only** powertrain (R10). Battery lean **4S LiPo** class (~14.8 V / ~16.8 V) — not a pack SKU (R11). Hip + knee actuators **TBD**, class research only (R12). Wheels **~4–6"** skinny rubber; **~6"** was the working hypothesis at this date (R13). Later notes prefer **5"** and put **in-wheel BLDC FOC** on the rim.

### 2026-09-20 — #5 modes (consolidated here)

Four **manual** control modes **before** autonomy (R14):

1. **`PARKED`** — safe idle / failsafe; no balance loop driving wheels
2. **`TWO_WHEEL`** — both wheels active; bipedal teleop
3. **`LEFT_ONLY`** — balance / drive on the left planted wheel; right leg free
4. **`RIGHT_ONLY`** — mirror

Pilot selects via **TBS Nano** (likely aux / flight-modes style). Wi‑Fi telem reports the active mode. Mode change does not lift, plant, or path-follow.

### 2026-09-20 — #6 CoG / hip-roll / reuse-control (consolidated here)

V1 overall width **~10"** (R15). One-leg balance is a **full loop** (R17): hip roll keeps weight over the planted wheel **and** that wheel drives fore/aft (inverted-pendulum pitch). **Reuse existing control** — do not write a novel Hux balance stack for V1 (R18). Pointers: XRobots TallBalancer / SonicRobot / RobotX, Hattori, Mini-Cheetah-style stacks, FC attitude loops, ODrive / FOC torque modes.

Hip roll was "TBD if mandatory for V1" in this packet. **#10 supersedes that softness:** hip roll is **in V1**.

### 2026-09-20 — #7 fabrication / 2D-before-3D (consolidated here)

- **R19** — Prefer cheap **COTS** stock for primary structure (carbon rods / tubes, metal stock, fasteners). Printed parts are joints / hubs / brackets.
- **R20** — Customs are **draft-friendly** (print now, mold later).
- **R21** — **Wire openings and ports** through links and body.
- **R22** — **Serviceable V1**: fasteners, pack, FC, actuators as replaceable modules.
- **R23** — **Several 2D sketch layouts before any Blender / 3D CAD.** Hard process gate.

### 2026-09-20 — #8 mass / wheel motors (consolidated here; mass hardness superseded)

Wheel motors sized for **reaction speed / torque bandwidth** (balance), not max continuous power (R25). Candidate *classes*: lightweight gimbal BLDC ~2208–4108 + FOC + encoder; mid small-outrunner + reduction if needed. Avoid SonicRobot-class 63xx / hoverboard hubs.

Original **under 6 lb** / aspirational **4–5 lb** target (R24) is kept as a **historical preference only**. **#15 supersedes** any hard mass ceiling: budget is **soft / blown**.

### 2026-09-20 — #9 leg actuators by axis (consolidated here)

Jobs differ — do not force one actuator type on every axis (R26). Trade write-up: [`research/actuators-legs.md`](research/actuators-legs.md).

| Axis | Job | Class lean (later refined) |
| --- | --- | --- |
| Hip roll | Highest bandwidth, continuous small corrections, prefer backdrivable | Dynamic FOC / QDD / fast servo — **locked in V1 by #10** |
| Hip swing | Position + speed for the step cycle | **Servo vs stepper+belt TBD** (#14) |
| Knee | Highest gravity + step torque for ~9.5" | **Servo vs stepper+belt TBD** (#14); springs / linkage still recommended |

### 2026-09-20 — #10 V1 actuator baseline + 8-axis I/O + hip roll in V1 (consolidated here; swing/knee lock lifted)

**Hip roll is IN V1** (R16 / R27): dynamic FOC BLDC / small QDD / fast bus servo. Experimental — may not work as hoped. Still ship the joint and `LEFT_ONLY` / `RIGHT_ONLY`. **Not a stepper. Not V2.**

This packet also locked knee + hip swing to **stepper + belt/gear**. **#14 lifts that lock.** What remains:

- Wheels stay **brushless FOC** (R6).
- **8-axis I/O** (R29) is the architecture **if** pose joints are steppers: 2 wheel BLDC + 4 steppers + 2 hip-roll dynamic.
- **FC does not drive stepper coils.** TMC-class / multi-axis driver board(s) between host and motors. Preferred: FC = IMU + wheel FOC (+ hip-roll if PWM/CAN); **Pi or dedicated stepper controller** = 4× step/dir.
- Drone firmware as a stepper host is a V1 **anti-pattern**.

### 2026-09-20 — #11 electronics-minimum (consolidated here)

Minimum electronics plan: [`electronics-minimum.md`](electronics-minimum.md). Text block diagram, parts *classes* (not SKUs), phased assemble order **P0–P5**. FC stays TBD. No spend.

### 2026-09-20 — #12 mechanical layout (consolidated here)

- **R30** — Wheel drive **BLDC at the wheel** (hub / coaxial), not remote from the hip.
- **R31** — **kV match** is a sizing *goal* (4S + diameter + balance bandwidth). No invented number.
- **R32** — If pose joints are steppers: mount **high** (above the knee); **belts** to the pivots.
- **R33** — If belts: **one inside, one outside**; integrate toothed pulley into the printed custom where draft-friendly.

Hip roll stays the V1 **dynamic** actuator. Placement TBD. Do not drop CoG with heavy roll actuators if avoidable.

### 2026-09-20 — #13 parts-on-hand + shop (consolidated here)

Inventory: [`parts-on-hand.md`](parts-on-hand.md). Shop tools (not parts): [`capabilities.md`](capabilities.md).

- **Zantle 5"** walker wheels ordered (ASIN B0D534PDRT). **OK to hack apart** — disposable donor rubber, not a part to preserve.
- Wheel diameter soft-preference moves to **5"** (still 4–6" band).
- Shop: mill, lathe, metal bender, metal brake, bandsaw, soldering, welding, breadboards. Hardware fab is intentional.

### 2026-09-20 — #14 Serra / Build Some Stuff + servo vs stepper TBD + 2× plant + 4S step-down (consolidated here)

Steal [Kelton Serra / Build Some Stuff](research/inspiration.md) **packaging**, not files: in-wheel BLDC + encoder; jointed legs keep CoG over contact as height changes; serviceable modular prints; wheel-under-CoG correction geometry.

- **R12 lifted:** knee / hip swing are **servo vs stepper+belt TBD**. No lean either way. Serra's 40 kg-class servos are a data point, not a Hux SKU.
- **R36:** plant-side knee / hip / hip roll see roughly **~2×** two-wheel stance. Size for that case.
- **R11 addendum:** keep **4S**; add **controlled step-down** (BEC / regulator *class*) to 5 V / 6 V / 7.4 V pose or logic rails so wheel FOC spikes do not brown out planted joints.

### 2026-09-21 — #15 carbon tubes + ~24" height + soft mass (consolidated here)

- **R34** — Primary **upper + lower leg spars** are **carbon fiber tubes**. Printed / machined **end fittings** only.
- **R35** — V1 height up to **~24" at full extension**. Width stays **~10"**. Still must reach **~9.5"** with margin.
- **R24** — Mass budget is **explicitly blown / soft**. Capability and packaging beat the old 4–5 lb / under 6 lb numbers. Those stay as historical preference, **not a kill-switch**.
- Longer tubes → longer belt runs if a belt is the reducer. Joint actuators sit at hip / knee with tube between.
- **GIM8108-class** noted as a knee / swing *candidate* (not ordered, not locked). Aligns with #14's open class.

### 2026-09-21 — consolidation PR (this file)

Open PRs **#5–#15** conflicted with each other and with `main` (which already had #1–#4). Unique content is merged here. **Latest intent wins** on actuators and mass. After this lands, close #5–#15 as obsolete.

### 2026-09-21 — Tazer lessons (folded into consolidation)

Steve shared [Tazer — My Robot almost got me Kicked out of Uni](https://www.youtube.com/watch?v=gqnW9qBCHnM): **learn a lot from this guy's mistakes.** Note: [`research/tazer-lessons.md`](research/tazer-lessons.md).

Does **not** change locks. Reinforces: GIM8108-class is a live *candidate* (still not ordered); carbon tubes confirmed; 4S stays preferred (lighter than his ~48 V) but FOC stalls still need a real power bus, not a logic PCB; start with simple PID / existing control (R18), not LQR-on-a-bad-model; rubber not TPU; carbon-dust PPE; serviceability + software torque limits later.

### 2026-09-21 — #17 Stompy CAD→sim→real

Steve asked to analyze [Stompy](https://www.youtube.com/watch?v=gEjg179fvmc) (Kayden Knapik — week-build RL walking biped), especially simulations and matching CAD to reality. Note: [`research/stompy-sim2real.md`](research/stompy-sim2real.md). Indexed from [`research/inspiration.md`](research/inspiration.md), [`research/study-plan.md`](research/study-plan.md), and [`research/README.md`](research/README.md).

Does **not** change locks. Walking ≠ Hux wheeled balance — still steal fixture / home pose, measure-vs-CAD, tether, default angles in CAD+firmware, and a later CAD→model lockstep for geometry / stairs. **Do not require RL walking for Hux V1.** Keep **reuse simple balance control** (R18). Sim is not a day-one wheel-balance task. No spend. No Jetson / Robstride / mjlab lock.

### 2026-09-21 — #18 Diablo wheeled-leg research

Steve shared [ETA Prime's Diablo review](https://www.youtube.com/watch?v=S5PoZ8aNwvs) plus paper [arXiv:2407.21500](https://ar5iv.labs.arxiv.org/html/2407.21500) (Direct Drive Tech / DDTRobot commercial self-balancing wheeled-leg). Note: [`research/diablo.md`](research/diablo.md). Indexed from [`research/inspiration.md`](research/inspiration.md), [`research/study-plan.md`](research/study-plan.md), and [`research/README.md`](research/README.md). Shop / SDK cited only.

Does **not** change locks. Steal: split brain (Pi vs motor board), DD/QDD as a *class* for high-bandwidth joints, **LQR/PID before RL**, height modes as states, aux contact later, payload vs height. **Do not buy Diablo.** Do not scale Hux to ~22 kg. No head tilt / cargo / creep rollers for V1. Keep **reuse simple balance** (R18). Hardware stays **TBD**. No spend. No M1502D / ROS2 / Pi4-as-FC lock.

### 2026-09-21 — leg geometry + wheel contact

Steve asked for a plan review and new leg math. The ordered walker wheels are not a good enough foot. He is considering a small spoked bike tire with real rubber. Carpet and interior floors can wait.

Study: [`research/leg-geometry.md`](research/leg-geometry.md). **No spend. No tire SKU. Spokes not locked.**

- **R13 updated.** The **4–6" / 5" preferred** contact plan is dropped. Zantle stays a disposable **bench donor**. Working draw is **~8" OD**, **~1–1.25" wide**, real rubber, torsionally stiff, until a real tread is measured. A **12–16"** kids bike wheel fails the tread, the **~10"** width, or the reaction-speed trade.
- Hip over the wheel: knee gravity is the poke (a few N·m); a spring cancels the two-leg share. Hip roll sees the large moment when the other wheel unloads (`m g ×` half-track, ~6.5 N·m at 6 kg on the draw).
- Tubes stay stiff (bending above ~40 Hz). Give lives in the spring and in radial tire compliance.
- The old **~0.8 N·m** wheel figure is historical (2 kg). Wheel target on the 8" draw is about **3–4 N·m peak**. Not a SKU.

### 2026-09-21 — wheel settled at 6"

Steve: the wheel has to fit comfortably on one stair so the robot can pivot and place the raised wheel on the next step. Settle the diameter early.

Study update: [`research/leg-geometry.md`](research/leg-geometry.md). **No spend. No tire SKU.**

- **Design step is 9.5" rise × 9.5" going**, nosing to nosing. The going matches the riser already in R3. A deeper measured tread is spare margin, not a reason to grow the wheel. Reopen only if a measured going is under ~9".
- **R13 settled at 6" overall diameter** (5.75–6.25" still counts), width ~1–1.25", real rubber, torsionally stiff. The whole tire sits between the nosing planes with about **±1.75"** of balance roll, and about **2.5"** of air under a 1" soffit.
- The **~8" working draw is withdrawn.** On a 9.5" going it leaves about ±0.75" of roll. A 12" kids wheel does not enter the slot.
- Leg draw follows: **7.5" + 7.5"** tubes, **6"** of body above the hip. Wheel peak about **3 N·m** on the 6 kg example. Not a motor SKU.

### 2026-09-21 — first buy

Steve asked whether a BOM existed and said he wants parts coming. There was no buy list. [`bom.md`](bom.md) is that list.

Spend is open **only** for 3× **6×1.25** ribbed pneumatic tires, 3× matching tubes, and 2× **1 m** carbon tube (**16 mm OD**, 12–14 mm ID). Motors, GIM8108, drivers, and another FC stay unauthorized. Zantle, the on-hand FC pile, TBS Nano, ESP32, and the Pi are not reordered.

### 2026-09-21 — knee to the rear, face, cameras

On the living drawings, Steve wants the **knee behind** the hip. Poke stays about **2.9"** at the 92% stance. The page also carries rough motor bulk (not a buy), seven cameras (front stereo pair plus back, sides, top, bottom), a small front display for preset faces, and an RGB in each eye socket. The lit socket is the eye. None of the face or camera parts are authorized to buy.

### 2026-09-21 — 14" wide, BOM prices

Overall width moves from **~10"** to **~14"** (R15). The head stays a narrower unit, about **7"** on the drawing. Wheels and legs are outside it. Track is about **12.75"** with a 1.25" tire, so the one-leg moment at 6 kg is about **9.5 N·m**.

[`bom.md`](bom.md) now has store links, line prices, and totals. **$89.41** is the authorized order. The working total of about **$1,140** includes class estimates for parts that are not chosen and not authorized.

### 2026-09-22 — stair climb dynamics in the living drawings

Steve asked for judgement calls on gravity, weight, CoM, momentum, inertia and motion in the stair animation, and for the tool to be improved. The climb in `tools/living-drawings/kin.js` now models the front wheel rolling back under the mass during the flight and catching the leftover after the crest, joint torques for both legs, the lateral sway and roll moment (with a front view), real-time playback, and design knobs that rebuild the step. Findings in [`research/stair-climb-dynamics.md`](research/stair-climb-dynamics.md): on the settled geometry the step-to gait is a **±6% precision throw** because the rear leg leaves the mass 2.7" behind the front contact; a **forward landing error of ½" makes the step impossible**, so aim at the rear of the slot; body CoM **1–2" ahead of the hip axes** is the cheapest fix and makes the step routine; the knee is a **~10.6 N·m** holding joint standing up over the front wheel, not the 4.4 N·m stance figure. **Nothing locked.** No spend. The lump masses are still a picture, not a weighed robot.

### 2026-09-22 — Hux bot advisory + digital-twin horizon

Steve set the Hux-bot role and a long-term north star. **Docs only. No lock lifted. No spend. No SKU. No BOM cart rewrite.**

- **Role:** Hux bot is advisory / tracking / brainstorming / adversarial opinions. Not a large-task executor. The repo stays the plan SoT. Planning notes get updated; builds / CAD / firmware / spend wait for a specific Steve ask.
- **Current state:** stance **~14"**, **6"** foot locked, BOM started, living-drawings + stair-climb-dynamics and Diablo / Stompy / Tazer research are on `main`. Plan language should talk about that machine, not the early ~10" / soft-5" era.
- **Inventory ≠ design driver.** Parts on hand inform options. Prefer the correct actuators and wheels over the shelf. Zantle 5" stays a bench donor (already documented).
- **Spend rails:** avoid wrong actuators / wheels; keep cheap. Authorized spend is whatever [`bom.md`](bom.md) already says. Do not expand the order-now cart. Steve owns those breakout edits.
- **Horizon:** identical digital twin + training dojo (ML/RL) so a policy trained in sim runs locally. Later domains: stairs, rubble, dirt, fall leaves, wet mud. **Phased:** V1 classical / reused balance + modes (R14 / R18) → later CAD/URDF twin lockstep (Stompy) → later dojo / RL for hard terrains. Do **not** replace R18 with an RL gate. Consistent with Tazer / Diablo / Stompy: LQR / PID before RL.
- **Motion control:** high-bandwidth FOC / QDD / model-based loops where they matter (wheels, hip roll). Class only. No vendor lock.

### 2026-09-22 — 3D sandbox in the living drawings

Steve asked for a 3D view of the robot in the living drawings, three.js with simple controls, "almost like a video game" but with accurate physics and kinematics, as a first pass at fleshing out the design. Added [`tools/living-drawings/sim.html`](../tools/living-drawings/sim.html): Rapier rigid bodies at 2 kHz built from the same `kin.js` geometry and lumps, torque-limited joints, an LQR balance loop at 500 Hz, keyboard / gamepad / touch driving, and a course with the 9.5" × 9.5" stair, ramps, sills, a curb and wet tile. `npm test` in that folder checks it headless. Findings in [`research/sim-sandbox.md`](research/sim-sandbox.md): a **1" sill** crosses only near 1 m/s (6" wheel needs μ ≈ 1.1 to climb it on traction; knee saturates on the hit); **PARKED has no rest pose** (it rolls onto its back over the knee housings); one-wheel hold is not solved in the sandbox yet; one-wheel hip-roll holding torque ~9 N·m agrees with the 2D number. **Flags:** this is a design toy, not the digital twin and not a V1 job ([`software.md`](software.md)); the 3D robot is built only from the 2D numbers, so it is not CAD ahead of R23. The rear **parking skid** in the sandbox is a **proposal, off by default, not a decision**. Nothing locked. No spend.

### 2026-09-23 — sandbox second pass: ride height, suspension, stumbles, one wheel

Steve asked for ride-height control with a **medium stance by default** (lower CoG), leg suspension with an active lift on impact, better stumble recovery, and a planned, coordinated one-leg stand (the free leg kicked out wildly). All in the 3D sandbox; findings in [`research/sim-sandbox.md`](research/sim-sandbox.md) (second pass). Default ride is now **75%** (hip 14.3" vs 16.8") with Low / Medium / High presets. Legs are a 3.5 Hz virtual spring-damper. An impact reflex and a stall hop lift the wheel over edges: the **1" sill now crosses from 0.5 to 1.0 m/s** (was only 1.0). A capture-point leg catch and a hip-roll side step raise sideways shove recovery from **3 to 5.5 N·s**. The one-leg kick was a controller bug (balance point ~4° off → hip roll slammed to its limit); LEFT / RIGHT ONLY is now a planned **poise** over one wheel with ~15% left on the other, holding ~3.3 N·m on the planted hip roll. **Free wheel fully off the floor is still not held**; experimental switch, off. **Flags:** 75% is a sandbox driving default, not a change to the 92% stance in `leg-geometry.md` / the stair climb; the suspension and hop need **torque-controlled, backdrivable knee and hip-swing joints (FOC / QDD class) or real springs**. That leans on the open "servo vs stepper+belt" row; **Steve's call**, not decided here. Nothing locked. No spend.

### 2026-09-23 — geometry / contact / feasibility honesty (283bd23)

Major honesty pass already on `main`, not previously logged here. Implemented in the living-drawings model and recorded in [`research/model-corrections.md`](research/model-corrections.md), [`research/sim-sandbox.md`](research/sim-sandbox.md), [`research/stair-climb-dynamics.md`](research/stair-climb-dynamics.md). **No lock change. No spend.**

Shared spatial FK/IK (side and front no longer imply different mechanisms); whole-robot CoM; tipping moment is not hip holding torque; stair playback no longer looks successful when it is not — **current stair candidate rejected** (197/201 spatial samples fail). Tire: 6" sphere → rounded convex **6 × 1.25"** hull (free-roll numerical error still ~3%). Joint stops, self-contact, delayed sensing / drive lag, yaw integral, observed support vs requested mode, refuse `PARKED` without rest support. Historical planar pack-forward / ±6% throw recommendations retired as design decisions. **True one-wheel hold still fails.** Collision and tire remain approximations — the visual cylinder and the rounded hull are not a motorcycle crown (see 2026-09-25).

### 2026-09-24 — SpdrBot Isaac Sim research

Steve shared [Indystry SpdrBot](https://www.youtube.com/watch?v=YDzHL2JSCHc) plus [Indystrycc/SpdrBot](https://github.com/Indystrycc/SpdrBot) (Isaac Sim / Isaac Lab 4-leg spider; Fusion → URDF → USD; RL then a hand-tuned gait). Note: [`research/spdrbot.md`](research/spdrbot.md). Indexed from [`research/inspiration.md`](research/inspiration.md), [`research/study-plan.md`](research/study-plan.md) (Phase E + research index), and [`research/README.md`](research/README.md).

Does **not** change locks. Docs only. No spend. No SKU. No BOM cart rewrite. No CAD / firmware. Steal for the **Phase E** twin / dojo horizon: observation parity with real sensors, CAD→URDF→USD lockstep, reward-hacking risk, validate in Sim before hardware, mid/zero before assembly, power delivery, friction hacks can hurt. **Do not** adopt spider morphology, the hobby-servo / Pico stack, buy Indystry packs, stand up Isaac Lab before `TWO_WHEEL`, assume RL drops onto the robot, or buy a 4090 for V1. Ties to Stompy CAD/reality, Diablo LQR-before-RL, Phase E in the study plan, and **R18** classical first.

### 2026-09-25 — twin pipeline TBD, hip-roll rethink open, tire crown, website sync

Steve asked to log the 2026-09-25 planning conversation while hardware and planning run in parallel. He wants pipeline *direction* before / while hardware. **Docs only. No lock lifted. No spend. No SKU. No BOM cart. No CAD / firmware.** Hip roll stays **IN V1** (R16). Twin / RL stays a **horizon**.

- **Twin / RL pipeline TBD.** Choosing a digital-twin / RL simulation software pipeline is open research, not a stack lock. **Lock the contract first:** CAD→URDF/USD or MJCF lockstep, observation parity with real sensors, motorcycle-crown tire contact, firmware zeros matching the model, and willingness to live in the stack. Do **not** stand up Isaac Lab, buy a 4090, or make RL a V1 balance gate before `TWO_WHEEL`. Do **not** lock Isaac / MuJoCo / mjlab. Matches SpdrBot / Diablo / Stompy / R18 / Phase E. See [`software.md`](software.md) and [`research/study-plan.md`](research/study-plan.md).
- **Living-drawings ≠ twin.** The 3D sandbox taught a lot and framed capabilities / expectations. It is a **design toy**, not the identical digital twin. Findings: [`research/sim-sandbox.md`](research/sim-sandbox.md).
- **Hip-roll rethink (open).** Steve does **not** like the hip-pivot unload to get onto one foot — it seems to work poorly. Log the dislike. **Do not remove the R16 lock.** Bad feel may be the wrong stair trajectory, the two-contact poise, or sim contact — not proof the DOF is useless. Separate (1) need some lateral CoG path, (2) is hip roll the cheapest DOF, (3) is the stair path asking an infeasible throw. The stair candidate is already **rejected** in [`research/model-corrections.md`](research/model-corrections.md) / [`research/stair-climb-dynamics.md`](research/stair-climb-dynamics.md).
- **Tire crown.** Sandbox wheels look like a flat ~90° edge profile, not a motorcycle-round rubber crown. A real Hux tire should keep a rubber contact pad when tipped / cambered. A hard edge messes lean / one-leg physics. Real world may be better; the twin **must** match the crown. Sep 23 already moved sphere → rounded convex 6×1.25 hull; visual is still `THREE.CylinderGeometry` (flat shoulders) and collision is not a true torus. Half-fixed. Twin / sandbox honesty requirement.
- **Website vs docs.** Living-drawings F765-Wing / Pi 5 pages are a concrete **bench plan** Steve wrote. Canonical docs still say **FC TBD** ([`electronics.md`](electronics.md), [`software.md`](software.md)). Keep website language from looking more locked than the plan.

### 2026-09-25 — tire crown and contact pad fixed in the sandbox

Steve: the tire profile / contact pad notes above "need to get fixed". Done in `tools/living-drawings/`: one shared cross-section (`tire.js`, full-round crown by default, `M.tireCrown` flattens it) now builds the Rapier hull, the Three.js lathe mesh and the 2D projection, so the flat-shoulder cylinder is gone and the contact walks the crown under camber (tested at 0–30°). Each wheel is a hub plus a **tread ring on a carcass spring** (`tireK` 40 kN/m, `tireZeta` 0.2 — guesses, knobs); the HUD shows squish and the implied pad. Side finding: the one-leg poise controller pivoted on the **hub**, but a cambered crown touches an inch away from under it — the poise was on its own bail-out threshold, and the compliant tire exposed it (fell). Pivoting on the crown contact (`contactOf`) fixed the poise (free-wheel load 11–31% vs 3–24%, planted hip roll 4.3 vs 6.6 N·m). Regression suite updated and passing. **No lock changed. No spend.** The tire stiffness is unmeasured; measure a real 6×1.25 before trusting the pad numbers. The twin-contract requirement (crown contact) is now met by the sandbox, which is still a design toy, not the twin. Findings: [`research/sim-sandbox.md`](research/sim-sandbox.md).

### 2026-09-26 — compute / stack review, then 6S → 8S, CAN, control core, ROS 2 companion

Steve asked for a review of the software stack and whether a "$400–600 NVIDIA SBC" fits ([`research/compute-stack-review.md`](research/compute-stack-review.md)). Findings: the F765-Wing has **no CAN**; every torque-mode actuator on the candidate list speaks CAN; the bench firmware plan lacked blackbox, live params, a framed link and a hardware torque cut; the Pi 5 is right for V1 and wrong for seven cameras; Jetson prices moved 2026-07-22 (Orin Nano Super kit $399, Orin NX 16 GB module $999).

Steve then said **nothing is locked**, he owns the Pi 5 and F765 but need not use them, and he wants to **make this a real project** — a stepping stone is fine if the logic carries, otherwise start on the better board. He approved **4S → 6S** (large pack or two in parallel; LiPo or similar).

**Decided (supersedes older language above):**

- **Power: 6S** class, ~22.2 V nominal / 25.2 V full. Later the same day Steve picked a pack — **one 6S 5200 mAh 60C LiPo with XT90** — then said it was picked somewhat randomly, a smaller pack is fine, he wants reasonable runtime, and **"add the anti spark"** (XT90-S on the harness — done). The actuator check that followed ([`research/actuator-shortlist.md`](research/actuator-shortlist.md)) found the RobStride 00/01/02 input floor is **24 V**, below a 6S pack for most of its discharge. **Steve, later the same day: "8S yes."** Pack is **one 8S 3300 mAh 50–60C LiPo, XT90**, XT90-S on the harness, in the order-now cart. Canonical pages swept 6S → 8S. Regulators must accept 36 V; the companion slot gets a 12–19 V buck. Rails: **5 V** (MCU, RX, Pi), **12–19 V** (companion slot; a Jetson kit needs this — 6S full exceeds its 19 V input), pose rail if servos. Motor bus is 6S direct.
- **Actuator bus is CAN.** Wheels, hip roll, and — working class — knee / hip swing are **CAN QDD / FOC actuators with their own PD** on 8S. Servo / stepper for knee and swing drop to **fallback** status; steppers stay off roll and wheels. **This bus decision picks the MCU.**
- **Control core is a portable library.** Estimator, mode machine and balance controller in plain C++, no hardware calls, unit-tested; the same code links into the MCU firmware, the companion and the twin. Carry-forward is by design, not by porting.
- **Real-time MCU with CAN** runs the core at 1 kHz with IMU, CRSF, hardware watchdog and torque cut. **F765-Wing is P0–P1 bench learning only** (CRSF, one SimpleFOC wheel). **Steve (same day): "add the CAN MCU to the project." Picked: Teensy 4.1** + ICM-42688-P breakout + 3× CAN transceivers, in the [`bom.md`](bom.md) order-now table (~$55). Teensy over H743-WING because eight classic-CAN nodes need at least two buses at 1 kHz.
- **Companion: ROS 2 on Linux.** "No ROS" (V1 minimalism) is withdrawn for the companion; the MCU stays plain. Pi 5 now, containerized, V4L2/GStreamer cameras, no Pi-specific libraries. **Jetson (Orin Nano Super kit class) is the P5 perception buy**, not a V1 buy; the head carries a companion slot sized for it.
- **First image** must include blackbox logging (SD / MCAP), live parameters, a framed CRC'd binary link before any setpoint goes down, encoder velocity in the estimator, and a hardware torque cut. "Blink and print" is not the first image.

**Not changed:** R14 modes before autonomy; R18 reuse an existing balance pattern, PID before LQR before RL; hip roll in V1; 6" wheel; 9.5" × 9.5" step; 2D before CAD; twin pipeline TBD, contract first; no spend beyond [`bom.md`](bom.md) until Steve asks. **Spend implied but not yet authorized:** a CAN RT board (~$30–70) and the actuators; Steve owns those cart lines.

---

### 2026-09-26 — one-leg stance study: shift mechanics, hold torque, why the stand fails, hop budget

Steve asked for more high-effort simulation on the physics and geometry of the one-leg movement: the one-leg stand needed much more work, the hip-roll weight-shift mechanics were unclear, and the one-leg action did not work in the sandbox. Findings in [`research/one-leg-stance.md`](research/one-leg-stance.md); closed form in `tools/living-drawings/frontal.js`, checked against the Rapier sandbox. Ran in parallel with the same-day stack decision above and is consistent with it (the hip-roll hold becomes a torque feed-forward the control core sends to a CAN QDD actuator's PD). **Docs and sandbox only. No lock changed. No spend.**

- **Shift mechanics.** Both wheels down, the legs + body + floor are a parallelogram: both hip rolls turn the same angle. **24.7°** puts the mass over one crown contact at the 92% stance (29.5° at 75%); the axle spacer costs, the crown walk helps. 9.7° of the ±34° roll travel is left.
- **Hold torque.** The planted hip roll carries the body + free leg cantilever whenever a wheel is up: **8.3 N·m continuous** at 5.4" hips (5.4 at 3.5", 3.1 at 2"), **12–13 N·m peak** through a hop. That is above a GIM8108-class nominal (7.5). The 2026-09-22 "0 if the sway is done first" was the tipping moment, not the joint torque.
- **The static one-wheel stand is not a controller problem.** It is an acrobot (passive crown contact, two hip rolls as the only actuators) with the body CoM at hip-axis height, so rolling the body barely moves the mass: a 10 mm CoM error costs a 40° body swing on the planted hip (9° with both hips, but 52° of free-leg swing). **Capture region 2–3 mm.** The sandbox confirms it even with roll limits, torque and delay removed. Not a V1 capability on this geometry, whatever the firmware.
- **Dynamic single support is what the stair needs**, and it works within a clock: from a ~8% poise, ≤ 0.3 s in the air with a ±20 mm CoM estimate (≤ 0.5 s with ±5 mm or a one-shot 10° hip swing, worth ~11 mm). The sandbox lands and returns a 0.2 s hop (asserted in `npm test`); longer hops land but the return is unfinished controller work.
- **Sandbox fixes with hardware meaning:** hip-roll position hold needs integral / gravity feed-forward (a 90 N·m/rad hold sagged 4° = 40 mm of mass shift when the free wheel left); one stiff hip and one soft on the closed parallelogram; no leg-length levelling while the mass is off centre.

**Flags for Steve (contradictions with what is on file, not changed here):** R17 frames one-leg balance as a full-loop *gate before any stair cycle*; the physics says the achievable gate is a timed hop, not a hold. R36's "~2×" plant-side sizing is right for the knee/swing but the hip roll needs the 8.3 N·m hold + 13 N·m peak number, not a ratio. The 14" width / 5.4" hip offset is what sets that hold; bringing the roll axes inboard is the cheapest lever and is a 2D-layout question (R23), not a decision made here.

### 2026-09-26 — temporary actuator lock; masses and limits applied to the models

Steve: "Put a temporary decision lock on all of those motors. We're going to proceed with that set and do more calculations. I will validate more before I buy one to test with. Apply the new weights and other specifications to the 3D models and kinematic simulations. Push decisions to main."

**Locked (temporary):** 4× RobStride 02 (knee, hip roll), 2× RobStride 00 (hip swing), 2× RobStride 05 (wheel). Not ordered. New `tools/living-drawings/actuators.js` holds the vendor numbers (mass, rated / peak torque, ratio, no-load speed at 48 V scaled to the 8S bus, housing envelope, voltage window) and the lump-mass picture derived from them; `spatial.js` (torque limits = peak), `kin.js` (lumps, envelopes, wheel torque knob), `sim-core.js` (caps, per-joint torque-speed lines, rated torque reported beside every joint torque, hub mass) and `frontal.js` (lumps) all read it. Change the set there and every model follows.

**What the models say with the real masses** (findings, logged in [`research/sim-sandbox.md`](research/sim-sandbox.md), [`research/stair-climb-dynamics.md`](research/stair-climb-dynamics.md), [`research/actuator-shortlist.md`](research/actuator-shortlist.md)):

- **Mass picture 7.75 kg**, up from 6.0: body 4.35 (8S pack inside), hips 1.50, knees 0.46 each, wheels 0.49 each. 2.56 kg of it is actuator; 1.9 kg of that is in the legs and hips.
- **Hip roll hold** (frontal model, one wheel up, free leg hanging): **10.7 N·m at the drawn 5.4" hips, 8.0 at 4", 7.0 at 3.5", 6.0 at 3", 4.0 at 2".** RS02 is rated 7 (one retailer says 6). At 5.4" even an RS06 (11 rated) sits at its rating. **Requirement: hip roll axes at ≤ 3" from the centreline.** The 2D layout owns this; the head is 7" wide, so the roll actuators (3.1" housings) sit at or inside the head's faces — a packaging problem to solve, not to wish away.
- **Knee, standing up over the front wheel: 13.6 N·m static** (was 10.6 at 6 kg). RS02 peak 17 covers it; rated 7 does not, so this is a ~1 s, ~2× rated event every step. Gravity springs on the knees (R7) go from nice-to-have back to planned.
- **Stair throw:** with the actuator masses in the legs the rear-leg shove no longer crests at body CoM 0 — the 2D climb solver fails "reversed" (lift 0.90 vs need 1.03 kg·m²/s). **Body CoM +1" ahead of the hip axes** restores a window (margin 0.11, window 0.15); +1.5" gives 0.22. The 2026-09-22 pack-forward call, retired as a decision on spatial grounds, is back as a mass-placement fact. The candidate stays rejected on reach; this is a second, independent reason.
- **Wheel:** RS05 at 5.5 N·m peak, 31 rad/s no-load on 8S (2.4 m/s at the 6" wheel). The 3 N·m stair-catch knob is now 5.5; the catch is friction-limited before it is torque-limited.
- **In-wheel packaging:** the RS05 housing is 44 mm long; the tire is 31.75 mm wide with the axle 0.975" outboard of the leg plane. The actuator is wider than the wheel and reaches into the spacer. Hub drawing needed before any wheel order.
- **Sandbox:** physics rate 2 → 3 kHz (the heavier RS05 hub on the 0.2 kg tread ring chattered at 2 kHz; a solver artefact). Two controller changes the mass forced, both already prescribed in `one-leg-stance.md`: levelling frozen for the whole one-leg sequence (a 2 Hz limit cycle otherwise), and the lump model's gravity feed-forward on the planted hip through shift / poise / landing (the P+I hold sagged into the 10.7 N·m cantilever and fell). **The hop peak on the planted hip is now 17 N·m — the RS02 cap** — at the drawn hips. The 1" sill regressed to 1.0 m/s only (leg-spring / hop timing tuned for 6 kg). `npm test` passes.

**Not changed:** 8S; CAN; the four layers; R14 / R18; hip roll in V1; 6" wheel; 9.5" step; twin pipeline TBD. **Pushed to `origin/main`** at Steve's request.

### 2026-09-26 — exploratory concept art on a media page

Steve asked for starting visuals before accurate models, then for those pieces on the HTML site. `art/` holds concept renders and four image-to-mesh studies (whole robot, leg, wheel, torso). [`tools/living-drawings/media.html`](../tools/living-drawings/media.html) shows them. **Nothing locked.** Not to scale, not CAD, and not a pass through the 2D-before-Blender gate. The drawings and the sandbox are unchanged.


### 2026-09-27 — direction: finish it; one step is the finish line; T-REX is the scale reference, not the leg

Steve, after two days on the knee-linkage question: "I need to pick a direction. Their robot is for production sale. Mine is for R&D and personal use … maybe I should take an easier route and basically just match what they have done. Perhaps going up one step is enough for my needs. I want to actually finish this project so I should not make it too difficult or expensive for myself. I guess if we keep the knee motors at the knee we can have enough leg articulation to do the stairs which is a major achievement. My top flat-traversal speed should be a bit more than a normal human walking speed but not quite a jog … I would like the more capable legs if we can realistically achieve the type of controls to handle it."

**Locked (four calls, Steve picked each):**

1. **V1 finish line = one 9.5" step, 9/10 from a standstill.** One step exercises the whole mechanism (roll shift, timed single support, shove-and-catch, landing in the slot); a flight is that cycle repeated with the error held inside ±1.75" every time — V2, same hardware. North star unchanged.
2. **1.5 m/s top / 1.0 cruise.** RS05 no-load on 8S ≈ 296 rpm nominal / 264 at cutoff = 2.4 / 2.1 m/s at the 6" wheel with nothing left; holding 2 N·m of catch gives ~1.5 m/s fresh, ~1.3 near cutoff. Leg length and top speed are not a trade: taller CoM falls slower; what caps speed is wheel torque at rpm. The one coupling is bump impulse at speed (controls).
3. **Knee RS02 at the knee, spring in scope.** Simplest build; hip-driven linkage parked as V2. Five-bar rejected on the numbers (`research/knee-linkage.md`, `tools/living-drawings/studies/fivebar-check.py`).
4. **V1 terrain = flat + 1" sills + ~20° slopes.** Rough ground is V3 / Phase E.

**Why "match T-REX" was not the easy route:** matching it means a 150 mm-stroke five-bar and the step is gone for good on that frame; the serial stair leg costs the same eight actuators and only design time on the hip carriage. T-REX is the scale and packaging reference (row above). Orin Nano confirms the P5 companion slot; V1 stays Pi 5 + Teensy.

**Not changed:** actuator lock, 8S, CAN, four layers, 6" wheel, 9.5" step, hip roll in V1, ≤ 3" roll axes, 2D before Blender. **Uncommitted on the Mac; Steve commits and pushes.**

**Cascade, same day:** the four calls are numbers in `tools/living-drawings/spec.js` and every page reads them; **Sheet 1 — V1 layout** (`sheet.html`) is drawn from the model files — roll axes 3.0", 9.1" hip band under the 7" head, leg plane 4.75", knee RS02 inboard at the knee, RS00 outboard at the hip, pack 1" forward, axle spacer 1.63", landing aimed 0.75" rear of the slot centre. `software.md`, `electronics.md`, `mechanical.md`, both checklists, `study-plan.md`, `README.md` carry R37–R40. **Finding from the sheet:** with the wheel planes 3.375" outboard of the 3.0" roll axes, the 24° shift needs **2.8" of leg-length difference** to keep the body level; the sandbox needed a geometric levelling feed-forward to settle the poise at all ([`research/one-leg-stance.md`](research/one-leg-stance.md), addendum). That is a control-core requirement now (`software.md`). The roll-axis requirement stands; it was never free.

### 2026-09-27 — Sheet 2: tubes, fittings, knee spring, hub, wire path

Steve: "go ahead and do sheet two and then push everything to main."

**Drawn** (`tools/living-drawings/sheet2.html`, from the same model files; numbers in `spec.js` → `sheet2`): proposals, not buys — everything is drawn to parts already in the order-now cart or the temporary actuator lock.

- **Tubes:** the BOM's 16 × 14 mm carbon for both links. Joint axis → tube end 35 / 45 / 30 mm at hip / knee / axle (half the housing plus a wall), 40 mm bonded + cross-pinned sockets: upper cut ~191 mm, lower ~196 mm, four cuts = 0.77 m of the 2 m on order. Knee bending 12.4 N·m (stand-up) → 75 MPa, 17 N·m peak → 102 MPa on a 166 mm³ section; the bonded socket is the limit, not the tube. Hub offset (41 mm): 3.1 N·m bending under the one-leg load, 3.0 N·m torsion from peak drive. First mode with the wheel on the exposed lower tube ~116 Hz (E 100 GPa assumed).
- **Knee spring:** an extension spring along the rear of the upper tube, a cable over a **Ø2.5" pulley on the knee arm** (a two-anchor spring across the joint reverses past ~2·atan(offset/anchor) of fold). Fitted to the two-leg gravity torque at 92% (2.8 N·m) and 75% (4.8): **3.05 N·m/rad + 0.4 N·m preload → 3.0 kN/m (17 lbf/in) spring, 12 N preload, 3.4" travel to the 155° stop, 271 N there.** Result: **stand-up hold 12.4 → 6.4 N·m at the motor, under the RS02's 7 rated**; the swing leg holds ~5.6 N·m against the spring while raised. A torsion spring on the knee-arm boss or a gas spring are the alternatives; sized on the bench.
- **Hub:** the RS05 moves inboard so its outboard face is **flush with the tire** — nothing proud, nothing past the 14" envelope (closes Sheet 1's open call 10). Stator on the axle fitting F5 inboard (0.52" from the leg plane), disc web from the output flange to a turned 3.75"-bead rim. **Open:** the rim rides cantilevered on the RS05's output bearing (38 / 76 / ~230 N static / one-leg / landing) — rating not in the spec table.
- **Hip band:** one wider lower shell (9.1" × 3.1" × ~3.0", chamfered to the 7" head), not pods (closes open call 11). Pack, XT90-S and step-down on its floor; Teensy + IMU above. Yokes roll on the flanges outside the shell, so no slot.
- **Wires (R21):** bundles inside both tubes, ports in F2 / F3 / F5, service loops at the knee and hip, entry into the band's lower front face beside each roll housing. No belts, so R33 does not apply.
- **Fittings F1–F7** listed on the sheet with process (F1 / F5 / F6 machined, F2–F4 / F7 printed with inserts, draft-friendly). Bolt patterns from the STEP files in `cad/vendor` — not drawn.

**R23 status:** Sheet 1 (layout) and Sheet 2 (make-up) exist for the leg and the hip band. Blender can open for the head's styling; the leg and band go to CAD from these two sheets once Steve agrees them. Pushed to `main` with everything from 2026-09-27.


### 2026-09-28 — head audit and stair architecture reopened

Steve requested a high-effort pass focused on math, physics, geometry, kinematics and BOM, explicitly said **no components have been bought**, and authorized better component choices. During the review he identified the fundamental failure of single-leg support and asked that the legs justify the actuator cost rather than quietly abandoning stairs.

Steve subsequently instructed: **“Commit and push to main.”** This authorizes publishing the engineering review and its reproducible analyses; it does not change the validation gates or constitute approval to purchase components or fabricate the candidate.

**Firm corrections:** no confirmed Hux purchases/orders; old motor set is a reference, not a release; a floor hop does not qualify R2/R17/R37; battery location is not body CoM; stationary torque and actual cooling must govern holds; the old two-bus 1 kHz claim does not fit extended classic CAN. Added R41–R44. 8S remains conservative because July manuals say 24 V minimum while September tables say 15 V; hardware revision must resolve that conflict.

**Candidate H1:** H1 revision B: 203.2 mm depth, 120 mm middle bay, 177.8 mm top cap above Z = 80 mm; top +100 mm, cassette bottom −45 mm. The 242 mm-wide motor cassette is forward at X 32…89 mm; motor centers X = +60.5 mm. This fixes rear-folding link/motor interference and clears the rolled hip-pitch housings; the previous aft cassette/full-width middle bay are rejected. 150 × 50 × 60 mm pack class and separated compute/control/power allocations. Camera space is included; the control bay is 60 × 90 × 22 mm at X = 50.5 mm. Cassette width includes a 2.5 mm wall and 3.05 mm clearance outside each motor. Itemized nominal head 2.26 kg at (+6.19, 0, +20.78) mm; high case +0.60 kg. These are estimates, including explicit contingency, not measurements or a fabrication release.

**Leg results:** revision B's split-offset spatial model screens a complete step at 91 poses. The 24-inch wide-entry case fails 14 reach samples. Narrowing the stepping track to 120 mm reduces the lateral cantilever, but the 24-inch variant still fails three reach samples and flags 17 envelopes; the 25-inch variant hits the conservative knee/riser envelope during transfer. With 9.5-inch links / 658.8 mm (25.94-inch) height, the narrow-entry candidate connects and passes 810 interpolated reach/limited-clearance/width checks in both nominal and +0.60 kg head scenarios. Modeled nominal COM residual is 0.17 mm; peak span 355.44 mm leaves only 0.16 mm below the nominal width limit. Gravity roll/pitch/knee demands are 6.61/3.92/7.89 N·m, or 7.02/3.84/8.95 N·m in the heavier case. **Carry this 26-inch candidate forward for research; no hardware release or approved height change.** The 27-inch variant also connects but adds height and knee demand. A wide-entry 27-inch variant spans 547.1 mm despite its 14-inch initial wheel envelope, so wheel track alone must never stand in for moving-body width.

**Support mechanism:** keep hip roll, add active ankle roll and finite-width contacts. The ten-axis study still requires a real passive pitch-level carrier; it is not designed. Simply relocating the hip motors to the ankles is not a drop-in fix. Positively locked deployable landing shoes remain an alternative to investigate.

**Actuator finding:** vendor stationary references are RS02 6 N·m, RS00 3.6, RS05 1.2, RS06 8, conditional on vendor fixtures. Compact candidate hip roll reaches about 9.29 N·m; the preferred narrow-entry candidate reduces it to 6.61 N·m before dynamics, but hip pitch and knee also exceed reference stationary ratings. Do not substitute a 7 N·m rotating rating or a 17 N·m peak to pass it. Reduction/larger actuators require new packaging and duty calculations. Full-set procurement stays on hold.

**Cascade:** `tools/engineering/baseline.json` and `bom.json` are candidate numerical sources; `review.py` regenerates `head-leg-results.json`, H1 dimensioned SVG, the engineering page and the BOM. Physical-invariant tests cover mass/inertia, CoM sensitivity, split-offset FK/IK, gravity virtual work, contact-load sequencing, actual moving width and negative clearance/CAN gates. The full findings and remaining checks are in [head-and-leg-review.md](head-and-leg-review.md). Requirements, mechanical, electronics, software interfaces, checklists, inventory, README and notes point to the new disposition. Legacy model geometry/controllers are deliberately not relabeled as the ten-axis robot; pages visibly identify them as the rejected old stair baseline.

**Budget:** complete candidate allowance $2,236–3,498 before tax/shipping, with no assumed free boards/charger/stock. Price allowances are unverified; no supplier contacted, no order made. Firmware and validated CAD do not yet exist. **No component selection or dimension is called build-ready by this review.**

When a docs PR merges, add a dated heading:

```
### YYYY-MM-DD — #<n> <short title> (merged)

One-paragraph intent. Note any ID it owns or supersedes.
```

Keep the **Current governing decisions** table honest if the new PR changes a lock.

### 2026-09-28 — V1-PROOF scope reset; stair planning parked

Steve: “This project is very ambitious and basically too expensive. I want to shift all our planning to the side and create a new V1-PROOF model. This V1-PROOF model should be under $1000 and use stuff we have around. We can use a smaller target mass, less actuators, drop the stair stepping, use smaller actuators.”

This supersedes the stair finish line and its hardware constraints for active work. Preserve the old plan in STAIR-V1 and keep its numerical model and simulator labeled as historical. V1-PROOF has its own requirements, numerical input, budget, calculations, 2D sketch and landing page.

Proposed implementation: 2.5 kg target / 3.0 kg maximum, two encoder wheel gearmotors plus two reduced small leg servos. First balance with the legs pinned; then add slow synchronized height adjustment while both wheels stay down. Roughly 100 mm wheels, 278–306 mm upright height, 255 mm outside width and 28.5 mm theoretical height adjustment. The four-bar also sweeps the axle 49.3 mm fore/aft, so pose-dependent measured CoM/pitch trim is required. No stair-upgrade promise.

Cost allocation: $715 parts and fixture + $100 tax/shipping + $125 repair/overrun reserve = $940. Exact reusable stock remains unconfirmed, so no free-inventory credits are taken. Manufacturer motor/driver/servo references were checked for pricing/specification plausibility; other rows are caps pending quotes, not a shopping cart. Existing tools and personal fabrication labor are assumed. Any new tooling or outsourcing must fit the same cap.

No old BLDC-at-rim, CAN, 8S, carbon-spar, large-head, four manual modes or onboard-compute lock applies. Use suitable available controllers/IMUs and a compatible lower-voltage pack. Finish-line trials cover two-wheel balance, slow manual driving, leg height, fault handling and a ten-minute session. All physical tests remain open. No components bought, firmware flashed, physical robot built or site deployed by this planning revision.

**Publication:** Steve subsequently requested “push to main,” authorizing this V1-PROOF planning/model revision to be committed and pushed directly to `main`. Seven active model checks and generated-output consistency checks pass. Pre-existing untracked vendor CAD and downloaded mesh folders are outside this revision.

### 2026-09-28 — V1-PROOF 3D sandbox

User asked for V1-PROOF in the same 3D sandbox as before, to play with it and understand its movement and size. Added [`proof-sandbox.html`](../tools/living-drawings/proof-sandbox.html): the browser loads the study's own `tools/v1-proof/sim.js` (now also loadable as a browser script; plant and controller unchanged), so the driving feel is the controller that produced the 388/391 evidence, not a look-alike. The course lanes are the physical protocol's fixtures plus the 20 mm challenge threshold; shoves are the protocol's 0.8 / 0.4 N·s pulses and the 4 N·s failure case. Size references: floor grid in 10 cm squares, 12 oz can, US Letter sheet and the parked STAIR-V1 envelope (610 × 356 mm). The old `sim.html` stays as the parked stair sandbox.

The full study was rerun after the wrapper change: every pass/fail is identical and numbers match the saved results to 2 × 10⁻¹³; only the `sim.js` hash changed. **Flags:** the leg-height slider is a quasi-static preview (CoM and trim moved at 10°/s; no servo dynamics or reaction), and the fixtures have ramp kinks the study's tilted floor did not. The sandbox draws the body at a live pitch trim of about ±10° at the 15°/45° leg extremes. That tilt makes the body corner about 317 mm high at the tall pose, versus 306 mm upright. The motor/link packaging near the axle is drawn schematically. No hardware, purchases or new evidence.

### 2026-09-28 — V1-PROOF mobility and hardware baseline

User requested additional simulations, reliable movement in all directions and in-place rotation, push/uneven-surface tolerance, hardware decisions and a sound progression to later versions. Added an independent four-axis-proof simulation with the legs pinned, contact physics, actuator/sensor uncertainty, quantitative maneuver gates, saved failures and timestep checks. Added a hardware decision record and measured acceptance protocol; selected the smaller DRV8874 driver and protected 9 V servo branch. Budget changes from $940 to $900 including unchanged $225 reserves. No hardware tests, purchases, commit, merge or deployment are claimed by this entry.

Final evidence for revision B: 388/391 in the unchanged broad uncertainty matrix; all six directional maneuver cases pass 17/17 configurations. Retain three settling failures at the near −4 mm CoM corner and require ±2 mm fore/aft mass placement before grade/disturbance qualification; six separate adjusted-placement checks pass. Higher derivative gain was rejected for excessive modeled RMS current under delay/lost motion. Physics/controller regression checks and timestep consistency pass; no physical acceptance box is closed.

**Publication:** User subsequently requested “Push to remote main,” authorizing this simulation, hardware baseline and acceptance protocol to be committed and pushed directly to `main`. Pre-existing untracked vendor CAD and downloaded mesh folders are outside this revision.