Body next plan (milestone M5, stages P2-P4): after the power skid and the open frame

Written 2026-10-27 by the body planner (Fable 5.1) for the user and for the Opus implementer who will work it. Sources read, all on main 0e1b0ec: docs/body-plan.md (the plan P0/P1 were built from; its part list is stale), docs/body-frame.md, docs/body-electrical.md, docs/body-safety.md, docs/body-orders.md, cad/body_*.py, cad/build_body.py, cad/test_body.py, exports/body/{body,bom,mass,leg-loads,wiring,print-plan}.json, exports/body/fea/summary.json, exports/body/sendcutsend/, catalog/body.json, docs/decisions.yaml (D39, D40, D43-D47, D54-D56), docs/turret-next-plan.md (sections 3-5: the proposed D48-D52), exports/turret/turret.json, exports/leg/{leg,rig-results}.json, exports/leg/fea/summary.json, cad/fivebar.py frame_mounts(), sim/leg_rig.py, sim/robot_v2.py, exports/sim/v2/summary.json, the leg/stops and legrig/program branches (read only), docs/requirements.md, docs/milestones.yaml, docs/head-payload.md, docs/handoff.md, and the J501 project (~/Code/j501-agx cad/parameters.json revision 0.9-g48r, exports/simulation-checks.json; read only). SendCutSend's 5052 and 6061 material pages were read today (section 2.3). The arithmetic behind section 2 is in /mnt/storage-10tb/claude-scratch/gyropup/body-planner/ledger_levers.py (its output beside it).

Every "decide" below is decided here (D41's rule: the user does not wait on decisions); they are written up as proposed decisions D57-D64 in section 6 for the implementer to record. Task numbers in this plan are B0-B17; the old plan's T-numbers are quoted where a task continues one of them.

0. In one paragraph

P0 (the power skid) and P1 (the open frame) are a sound build package: seven SendCutSend parts that pass every published rule, a frame at SF 2.8-4.4 in the leg's real landing loads, zero clashes over 22 stances, a 13S bus whose every part is rated, and an order list where every line has a price, a link and a backup. Order it today with three small fixes (section 1.2). The one thing the plan got wrong is the mass: the body and head came out at 13.3 kg against the 10.3 kg budget and the robot at 32.3 kg against 28.5, mostly for good reasons (3/16 in cheeks because 1/8 in failed, a 1/8 in tray, a power plate and a printed battery bay the plan did not have, the turret's 1.70 kg, the legs' thicker flanges and machined adapters). That budget cannot be met by any body that holds these parts, so it is retired. The call (section 2.4): the robot is designed for 32.5 kg (16.25 kg a leg) in every strength case, held under 32.0 kg as built by a ledger that gates every commit, and expected at about 31.8 kg after the levers of section 2.3 (about -0.7 kg scheduled for P2/P3, -1.0 kg more on call). The legs must be sized for 16.25 kg: at that mass their three weakest checks (the wheel link's sideways case 1.03, the front link 1.10, the hub dowel bearing 1.12) sit under the 1.2 target, so the leg workstream's stop redesign is the moment to fix them. Then P2 puts the turret on the deck over a 6061 mount plate that takes the tip-over case from SF 1.3 to over 2, P3 closes the shell in 2 mm PETG with a fuse board and an aluminium battery tub, and P4 is the dock. The sim, the firmware and the thermal check (the old T4, T7, T9, T10) are tasks B3, B6-B8 and B9 here.

1. Review: P0 and P1 as built

1.1 Ready to order and cut today

Everything the leg's interface asked for is taken (the six inserts, the plane, the harness exits past the cheeks' ends, the wheel harness through the cheek windows), and both of the turret's cross-workstream items (the lidar 143 mm ahead of the pan axis, the harness grommet at the right rear (-130, -60)) are in cad/body_params.py with tests.

1.2 Change before ordering

1. The live ends of the branch leads are sockets, not pins. The wire list puts an XT60H male on the body's end of every RS-04 and VESC branch (W10-W15) and an XT30 male on the 24 V and 12 V branches (W21, W24): those pins are live whenever the pack is in (the + bus is unswitched; the switch is on the negative). The same rule the charge port already follows (hazard 11: "no live pins on the robot") applies: the body's branch leads end in XT60H-F / XT30U-F (sockets); the motor pigtails, the VESC leads and the J501's and lidar box's adapter leads carry the males. In the Pyrodrone order: 2 x XT60H-F 5-pack instead of the XT60H-M 5-packs (the females for the charge port stay); XT30U-F 8-pack for the body, the XT30U-M 8-pack for the device leads. $0 net. cad/body_wiring.py W10-W15, W21, W24 and the leg's pigtail request (section 5.1 item 5) say which end is which.
2. Two DDR-120C-24, not one. Master Electronics flags it "no longer being manufactured" (docs/body-plan.md 1.3); the old plan's instruction to buy two stands. +$69.20 (Bravo Electro). The SD-100C-24 is the successor if both ever die.
3. Read the instant prices before the cart goes (the ~$383 is an estimate) and record them in exports/body/bom.json made_to_order with the date, so the Build page stops saying "estimate".

Nothing else changes before ordering. In particular the cheeks stay 3/16 in (section 2.3), the deck is cut as drawn (the turret mount plate of P2 bolts under it through its existing riser holes), and the PETG battery bay is accepted for the bench and the tether (procedure: charging attended, on a non-flammable surface) until P3's aluminium tub.

1.3 Risks carried into P2-P4

2. The mass problem, first

2.1 Where the 32.3 kg are (exports/body/mass.json, stage P3)

Against the plan's section 1.8 estimate (its table says 29.5 kg; its rows add to 30.2): the cheeks +0.5 (3/16 in on standoffs against a 1/8 in U), the tray +0.15, the deck +0.2, the power plate with its posts and bus bars +0.36 (the plan hung those parts on the belly plate), the printed battery bay +0.36 (the plan's cassette was inside the pack's 2.0 kg, and the pack itself came in 0.46 lighter), the shell +0.2, the G48R cooler and spreader +0.18 against the plan's 34 mm cooler, the turret +0.36, the legs +0.7 (D46's machined adapters and the flanges); the converters -0.18 and the battery -0.46 went the other way. About 16.8 kg of the 32.3 are bought parts the body cannot change (motors, wheels, lidar and cable, turret, compute, battery, VESCs, converters); the rest is the legs' printed parts (5.2), the frame (4.3), the shell (1.4) and wiring and hardware (0.7).

2.2 What every kilogram costs the legs

The leg rig's loads rise about in proportion to the mass it carries: from the leg's committed 14.75 kg a leg (exports/leg/rig-results.json) to the body's 15.97 kg (exports/body/leg-loads.json), +8.3 % of mass gave +7.1 % on the drop_150 axle force, +9.4 % on jump_peak and +8.5 % on the undamped power cut; the knee pins +6 to +10 %. (The jump's push-off is torque-capped, so it does not grow; the landings and the collapse onto the stops carry the mass.) The leg's strength checks (exports/leg/fea/summary.json, 14.75 kg) already have three under their 1.2 rare-event target or on it:

So the robot as the ledger stands already puts three leg checks under target, and the leg workstream's stop redesign (branch leg/stops) is the right moment to fix them for the mass the robot will really have, once, instead of twice. The gearbox torques (99-108 N m in the jumps at 15.97 kg) stay under the RS-04's 120, and the jump at the 60 N m cap drops from 268 mm (14.75 kg) to 258 mm (15.97): about 4 % per 8 % of mass.

2.3 The levers, ranked

Grams are against exports/body/mass.json; costs are today's prices from docs/body-orders.md where the part is already listed and estimates (est) elsewhere. SendCutSend's 5052-H32 comes in 0.040, 0.063, 0.080, 0.090, 0.100, 0.125, 0.187 and 0.250 in (bendable) and 6061-T6 in 0.040-0.750 in (not bent), read today: there is no 0.160 in, so the cheeks' only step down is 0.125 in, which failed the landing case.

Scheduled (B2, B4, B5, B6): about -0.7 kg

On call, in this order, when the ledger needs them (D57): about -1.0 kg more

Not pulled, and why

• The battery stays 13S2P P30B (1.54 kg). It is already the light choice (D54; the P50B brick is 2.17 kg). A 13S1P cannot deliver the 80 A jump pulse (36 A a cell) and would sag 18 V; a 13S1P of 21700 P45B (1.22 kg, -0.3 kg) sags 15.6 V at 80 A, under the DDR-120C's minimum from a half-charged pack. The pulse sets the pack: two strings of 36 A cells is the floor. The jump gate (section 1.3) is the mitigation for the sag the 2P pack already has.
• The J501 box is not on the robot (M5's done_when: "no separate box"); the ledger carries the carrier, module, spreader and cooler only (1.16 kg). The box's 0.4 kg of PLA was never counted.
• The head's parts: the VLP-16 (0.83) and the turret (1.70, D49) are fixed by their workstreams; the lidar box and coil (0.35) are the cable's.
• The riser stays solid (0.17 kg): it is the turret's load path into the deck and holds the TM inserts.
• The ear pods stay (0.31 kg): R27's forward lights and the character (the ears).
• The 3/16 in cheeks stay for P1 (section 1.2): 1/8 in was SF 0.6 in the landing; lever 11 is a study for later.

2.4 The call

• The design mass is 32.5 kg: 16.25 kg a leg in every strength case of the legs (cad/leg_studies.py, the rig at --half-mass 16.25, at BUS_V 48.1) and of the body (cad/body_legloads.py --half-mass 16.25, cad/body_studies.py). The body's cases at 15.97 kg move by +1.8 % in load: S1 4.42 to about 4.34 away from the standoff ends, S1F 2.88 to 2.83, everything still passes (B3 re-runs them).
• The ledger ceiling as built is 32.0 kg: legs 9.6 kg each (19.2: the 9.48 of today plus 0.12 for the stops' redesign and the three fixes above), body and head <= 12.8 kg. cad/body_layout.py carries these as the budget and cad/test_body.py fails any commit that puts the P3 ledger over them (B2). The plan's 28.5 / 10.3 kg budget is retired.
• The expected robot is about 31.8 kg: 32.27 as built, -0.7 kg of scheduled levers, +0.24 kg for the legs at 9.6 (31.6 if the legs stay at 9.48). If the ledger crosses 32.0 (the harness, hardware and cooler masses are estimates until P1 is weighed), the on-call levers are taken in order 10, 11, 12 before the next part is ordered.
• What the legs are designed for: 16.25 kg a leg at 48.1 V; the wheel link's R1w, the front link's R1 and the hub's dowel bearing brought to >= 1.2 at that mass (today 1.03, 1.10, 1.12 scaled); the hip frame's insert pull-out coupon (the leg plan's bench test B4) at >= 2.4 kN; the jump at the cap re-run (expect about 250 mm at 16.25 kg and 48.1 V). The leg's extra material for those fixes is inside the 0.12 kg a leg the ceiling allows.
• Why not lighter: 30 kg would need every lever including the cheeks in 1/8 in (unproven), the other cooler, a cut lidar cable and the leg's printed parts 10 % lighter; 30.6-30.8 kg is the floor with every lever pulled (the scratch script). Designing the legs for a number the robot cannot meet is how the 14.75 kg rig got 1.13 on the wheel link.

3. The stages

3.1 P2: the head (after P1 balances on the tether)

The turret mount plate under the deck (B4), the riser, the turret on the TM pattern (after the turret's N16 sign-off), the DT12 harness and the STOP contact (already wired in P0: W34-W37), the lidar; the turret's own N18 tests on the robot (the barrel and every ball 10 mm clear of the lidar at every pan, the 30 cm drop on the stand); then target practice tethered, then untethered in a padded room. P2 also fits the hollow posts and the 10 AWG harness if P1's parts are reprinted or re-cut for any other reason (otherwise they wait for P3). Blocked on: the turret's exports/turret/mount.json (N5) for the final plate and riser (the plug's sweep box), and the turret hardware (N12-N16).

3.2 P3: the shell, the fuse board, the tub, the thermal test, the hazard review

Side walls in 2 mm ribbed PETG with the ear pods, the door and the port hatches cut from the mated plugs; the front face (eyes, D435 cradle); the rear (D435, four RP-SMA bulkheads, the radio receiver bay and stop jack, the switch's 16 mm button, the charge port); the belly box skids; the knee wedge, apex fillet and stop fairings against the leg's stops as merged; the battery tub in bent 0.063 in 5052 with printed end caps (the PETG bay burns: hazard 18); the fuse board with the shunt and INA228 on it; the Rev B tray, deck and pallet if the ledger needs them. The thermal test at 40 W with a thermometer; the touch temperatures; the hazard review on the P3 robot with the clash list before any session with children.

3.3 P4: the dock

The old plan's section 2.5 with one change: the battery is in the belly now, so the charge pads sit in the tub's floor, the lowest point of the body, and the dock is 30 mm tall so the pads land with the legs 8 mm above their stops (B9).

4. Tasks for the Opus implementer

Estimates are agent working hours; "compute" is FEA or MuJoCo time; "bench" is days with hardware in hand, where the user's hands are needed (the implementer writes the procedure and the log). Work in a feat/ worktree of ~/Code/gyropup with a PR into dev. New CAD goes in cad/body_*.py, exports in exports/body/. Nothing in cad/fivebar.py, cad/turret.py or sim/leg_rig.py changes without the owning workstream (requests to them are section 5). Heavy results go to /mnt/storage-10tb/claude-scratch/gyropup/ (the root drive is nearly full); only the summaries are committed.

4.1 Now, without hardware

Order: B0, B1 (the user orders), B2, B3 and B4 together (B3's loads feed B4's D2), B5-B9 in any order, B10, B11 last. About 45 h of agent time and 8 h of compute: four or five sessions. Printing of the P1 parts and the leg's and turret's builds run alongside.

4.2 With hardware in hand

5. The interfaces to freeze, and what is blocked on them

5.1 With the leg workstream (cad/fivebar.py, sim/leg_rig.py, branch leg/stops)

1. The design mass: 16.25 kg a leg (D57). sim/leg_rig.py HALF_MASS 16.25 and BUS_V 48.1 (the body does not edit the rig; cad/body_legloads.py passes --half-mass); cad/leg_studies.py at those loads; the wheel link R1w, the front link R1 and the hub dowel bearing to >= 1.2 at that mass; the extra material <= 0.12 kg a leg; the jump re-run. The body's loads of docs/body-frame.md 3 were taken with the rig at 54 V: at 48.1 V the jump's speed cap is lower, so they are if anything conservative.
2. The hip-frame interface, frozen with the cheeks' order: six M5 inserts at (x, z) = (0, +-70), (+-83, +-70) on the plane y = -82.3, pilots Ø6.4 x 11 mm, the flange end faces flat on that plane (face [-95, 95] x [62, 77] and its mirror); M5 x 12 button heads at 3 N m from inside the gap. Any move of a boss or the plane after the cart is ordered is a new pair of cheeks ($120). The leg/stops branch does not touch them (checked: its leg.json and frame_mounts are unchanged).
3. Insert pull-out: the body's bottom row sees <= 0.7 kN a bolt in the landing, with the cheek bearing on the flange ends; the leg plan's bench test B4, the insert coupon (docs/leg-prototype-plan.md 5: pull-out >= 2.4 kN, torque against preload), is the body's evidence too. If it comes in under 1.4 kN the body's fallback is a through-bolted row (M5 nuts behind the flanges), agreed with the leg.
4. Harness exits, frozen: the RS-04 leads leave past the cheeks' ends (fivebar.harness_angle(): the front one at (127.6, 24.5) at +25 deg, the rear at (-131.9, -11.1) at -169 deg, exports/body/body.json) and turn into the gap; the wheel harness through the cheek window |x| <= 20, |z| <= 25; the zone between the motors' rear covers and the centre plane stays free (the gusset at x = 0, z 28..62 stays).
5. Motor pigtails: 450 mm from the rear cover, ending in XT60H-M (the body's branch lead carries the socket: section 1.2); the CAN on GH1.25 with 350 mm extensions to the MCU shelf; a 120 Ohm plug at the far motor of each leg.
6. The sweep envelope near the body, frozen as a test: upper links |y| 102.5..132.5 reaching z <= +31 within |x| <= 225; the knee eyes at |y| >= 83 (the belly plate is 116 wide for them: hazard 1); the stop pins on the frame's outer face only. The body's clash sweep (cad/build_body.py) runs on both sides' CI; when leg/stops merges, B6 re-runs it before the fairings are drawn.
7. The leg's mass: exports/leg/leg.json mass_g.total with the stops' redesign (today 9,282 g + 0.2 kg announced); the ledger reads it (B2). The ceiling allows 9.6 kg.

Blocked on the leg: the three strength fixes and the rig at 16.25 kg (B3's leg-side acceptance); leg/stops merged (B6's fairings); the coupon test (item 3); the pigtails (P1's wiring); the rig's bench test B1 (the RS-04's power-off damping, which sets the fault loads the body's S1F case uses: 275 N m undamped, 42 N m damped).

5.2 With the turret workstream (cad/turret.py, docs/turret-next-plan.md)

1. The TM face at (x, z) = (-40, 246) (deck 216 + riser 30), the lidar's axis 143 mm ahead at x 103, the foot Ø144 and the bumper Ø172 (its rear at x -126, 34 mm inside the deck's rear edge); cad/turret_lidar.py --x 143 is the turret's N5; body_params.LIDAR_TO_PAN is the one source.
2. The grommet at (-130, -60), right rear (D52 taken; test_turret_grommet_on_the_right_rear); the DT06 plug stands at y -78..-123 from the pan axis at z 265 and withdraws to -153: above the deck and the side wall's top, over the right ear pod (z 142..193). B6 keeps that box empty.
3. exports/turret/mount.json (N5): mass with and without balls, the centre, the plug's sweep box, the recoil; body_params.tm() already reads it and takes the larger of it and the 1.70 kg budget. Until it lands the riser and the plate are built to the budget and the plug box of docs/turret-next-plan.md 3.1.
4. The electrical interface, frozen as wired in P0: pins 1-2 +24 V through the stop relay's pole B and a 5 A MINI 58V fuse (W22, W34), 3-4 GND (W35), 5-6 CAN to the J501's can1 (W36; the J501 terminates), 7-8 the dry loop on pole A (W37; closed = run; opens on the radio stop, the receiver's failsafe and the sit). The DT12 is plugged only with the relay open (the turret's 3 x 1,000 uF). The 24 V rail is 120 W for the J501 and the turret together: the firmware staggers the flywheel start from the Jetson's boot; the DDR-240C-24 is the step up if B12 shows the rail at its limit.
5. The riser's TM inserts: M5 x 4, 8 mm deep in PA6-CF, under the plate of D59; the turret's 30 cm drop test on the stand (N18) is the turret's; the body's D2 case is the deck's.

Blocked on the turret: mount.json (B4's final geometry), the turret hardware signed off (N16) for P2, the N18 tests on the robot, and the lidar table at 143 mm in the turret's CONFIG.

5.3 With the J501 project (read only)

The carrier's holes, the module's tie-down pattern, the G48R cooler (102 x 97 x 50.3 over the spreader at PCB (88, 70.5) +-3 mm) and the ports come from ~/Code/j501-agx/cad/parameters.json revision 0.9-g48r, an uncommitted working tree on branch rev0.8-thermal. B10 snapshots it. The fan centre is measured from the PCB edges when the carrier is on the pallet (B14) and the snapshot corrected if it moved. The cooler tie-down keeps the user's rule (four M3 x 35 flat heads from below through the pallet into the module-to-heatsink standoffs, 4 mm of thread; the M3 x 35 is still to source: docs/body-orders.md 4).

6. Proposed decisions (for docs/decisions.yaml; the implementer records them in B11)

• D57 (proposed): the robot's design mass is 32.5 kg (16.25 kg a leg) in every strength case of the legs and the body; the ledger ceiling as built is 32.0 kg (legs 9.6 kg each, body and head 12.8 kg), gated by cad/test_body.py; the expected robot is about 31.8 kg. The plan's 28.5 / 10.3 kg budget is retired. Why: the body and head came out at 13.3 kg for reasons the FEA and the real parts forced (3/16 in cheeks where 1/8 in was SF 0.6, a 1/8 in tray, the power plate and the printed bay, the fitted G48R cooler, the turret's 1.70 kg, the legs' thicker flanges and machined adapters); 16.8 kg of the robot are bought parts the body cannot change; every lever in section 2.3 pulled reaches 30.6-30.8 kg, not 28.5. The legs' loads rise about in proportion to the mass (leg-loads.json against rig-results.json), and their three weakest checks are already under 1.2 at the ledger's mass, so the legs are sized once, for 16.25 kg, during the stop redesign. The levers are taken in order: scheduled (1-9) in P2/P3, on call (10-15) whenever the ledger crosses 32.0.
• D58 (proposed): P0 and P1 are ordered as built, with three fixes: every live end on the body is a socket (XT60H-F and XT30U-F on the branch leads; the device leads carry the males), two DDR-120C-24, and the SendCutSend instant prices recorded before the cart goes. The cheeks stay 3/16 in; the P1 tray, deck, pallet and PETG bay are not re-cut before P1 is weighed and tested. Why: the package passes every published rule and every strength case; a day's delay for pockets would buy 0.3 kg at the cost of re-running checked FEA on an unordered part; the socket rule is the charge port's rule (hazard 11) applied to eleven more plugs; the converter is flagged end-of-life by a distributor.
• D59 (proposed): the turret stands on the 30 mm printed riser over a 6061-T6 0.125 in mount plate (160 x 140, pocketed) bolted under the deck through the riser's feet and two PEM nuts, ordered with P2; the deck stays 0.100 in 5052; the turret tip-over case (D2) is held to SF >= 2.0 (the everyday target), not the 1.2 abuse target it passed at 1.31. Why: a 1.7 kg head 130 mm above the deck on a 0.100 in sheet is the body's lowest margin and the one a child's shove or a fall loads; two plates in bending with the 6061's 276 MPa take it over 2 for 135 g and about $25; the plate also stiffens the D1 hand case at the riser (0.51 mm maximum today).
• D60 (proposed): a JLCPCB fuse board (2 oz copper, twelve Keystone 3568 MINI holders, the 1 mOhm shunt with Kelvin pads and the INA228, the - star point and the frame-bond stud, 58 V markings) replaces the inline Bussmann HHM holders and the Blue Sea shunt in P3; the holders stay on P0-P2 labelled "58 V fuses only" (D55). Why: D55 named the board; it also removes 110 g of holders and leads and 85 g of shunt, puts every fuse where one glance finds it, and ends the 32 V holder on a 53 V bus.
• D61 (proposed): the shell is 2 mm ribbed PETG (PA6-CF at the ear-pod bosses), every port hatch cut from the mated plug with cadagent.connectors and proven by an insertion sweep, the battery tub a bent 0.063 in 5052 box with printed end caps, and the thermal acceptance is a thermometer on the robot (intake within 3 K of the room, plate <= 70 C, exhaust <= 45 C at 40 W) with the CFD run only after the J501 project's own Rev 0.9-g48r case is calibrated and steady. Why: the shell carries nothing (D44's frame does); 3 mm was the plan's guess and costs 240 g; the mating-plug rule is the cad-agent standard since the J501's windows sat 2 mm low; the PETG bay burns (hazard 18) and the 0.063 in tub is mass-neutral; the J501's CFD is not settled and a solver fighting the FEA on one machine proves nothing a thermometer cannot.
• D62 (proposed): sim/robot_v2.py models the body as built (the frame's solids, the tub as the lowest body, the ear pods, every mass from exports/body/mass.json, BUS_V 48.1, BODY_X +-160, Y_OFF from cad/body_params.py), the fifteen v2 scenarios are re-run at the ledger's mass, and the RL build regenerates its model from it while the trained policy stays frozen under GYROPUP_MODEL_DIR; the body's and the legs' strength loads come from the rig at the design mass, not the ledger. Why: the sim today balances a 26.2 kg robot with an interim shell; the controller gains and the leg studies must see 32 kg before P1 stands up; the RL policy is M6's and is not invalidated by a model it does not read.
• D63 (proposed): the balance firmware is C on the Nucleo-G474RE (three FDCAN, the VN-100 at 800 Hz, the 1 kHz loop from sim/controller_v2.py), with the RS-04 protocol as the leg rig's rig/rs04.py defines it and a Python HIL on the rig's emulator; it enforces UVLO (warn 45.5 V, sit 41.6 V), a jump gate (no jump under 44 V resting, height scaled with bus voltage), the Jetson heartbeat (2 s balance then sit), the STOP sense and the people-near caps; the Jetson decides "people near" from the lidar and both cameras (the lidar's rear 84 deg is behind the turret); the Daly BMS's discharge over-current delay is set to 3 s where its app allows. Why: D7 assumed the G474 and nothing better was named; the rig node's CircuitPython on a Feather M4 CAN serves one bus at the rig's rate, and the body needs three buses, the IMU and an FPU loop at 1 kHz; the 112 mOhm pack sags 9 V at 80 A, so a jump from an empty pack would take the 24 V rail under the DDR-120C's 33.6 V minimum; a 100 A BMS with a short delay would trip on a jump pulse.
• D64 (proposed): the dock is the plan's section 2.5 with the charge pads in the battery tub's floor and a 30 mm dock height: the robot reverses over an AprilTag on the rear D435 and the lidar, sits to a 108 mm leg length, two brass pads land in a printed funnel on spring contacts, a dock-present pair and the tag gate a 10 A charge relay in the charge port's line, the dock holds an HLG-320H-54A set to 53.3 V and 3 A, and the BMS does the rest. Why: D18 and R24 as decided; the tub is now the lowest point of the body (25 mm of clearance at the stops), so the pads go there and the dock must be lower than the plan's 40 mm; the charge port's fuse and the BMS already protect the path.

7. Still to verify, and TODO

• The SendCutSend instant prices (B1); the M3 x 35 flat heads (Belmetric's socket head is the fallback); the aluminium standoff for lever 12 (a McMaster number, read one page at a time); the Keystone 3568's current rating at 30 A (the VESC branches) against its datasheet; the 1 mOhm shunt's part number; Pololu's restock of the D42V110F12.
• The G48R's real mass (the ledger's 450 g is an estimate; weigh it in B14) and the DDR-120C's (450 g est against the datasheet's 510 g packed); the harness (400 g est) and hardware (300 g est) from P0's and P1's weigh-ins.
• The Flipsky card (the switch's side), the Daly card (the short-circuit trip and the over-current delay), the Luna charger's 90 % voltage metered at 53.3 V: all on arrival (B12).
• The leg's new leg.json total and the stops' positions after leg/stops merges; the turret's mount.json after N5.
• The cheek Rev B study (lever 11) needs P1's measured landing loads first: a strain gauge or the rig's drop test on the frame before any 1/8 in cheek is drawn.
• Not in this plan: the game page and the voice (M6), the RL policy at the new mass (M6), the leg's spring question (D39: after the damping test).

8. Notes from the work (2026-10-27)

• B8 (merged from body/b8-firmware): the LQR gain table in firmware/balance/include/gp/gp_model.h came from the interim 26.8 kg model; it is regenerated (firmware/balance/tools/gen_model.py) on B3's model of the real body, and again whenever sim/robot_v2.py changes (the 10S bus of D65 is the next). The jump-height scaling with bus voltage is the firmware agent's own rule (docs/body-firmware.md), not a sim result: B13 checks it on the bench. The motor calibrations in the firmware are placeholders until rig/calibrate.py runs on the bench (B13).
• The J501's fan blows up (j501-agx 94da39d): the body's louvres are the inlet and the deck's cells the outlet; the printed shroud became a duct (docs/body-thermal.md). The turret box (cad/turret_box.py) still assumes the fan draws in through the deck.
• D65 and D66 (main, 0b69b6d, 0a636a4): the 13S2P brick, its Daly BMS and its tub give way to two DeWalt DCB609 packs in 20 V mode in series (10S, 30-41 V) on two dock adapters; the DDR-120C-24 gives way to a converter rated to 28 V or below; the turret's 24 V branch gives way to an XT30 branch off the bus. The voltage windows above (UVLO 45.5 / 41.6 V, the 44 V jump gate) move to the 10S range: section 9 below records the new ones.

9. D65 and D66 in the body (2026-10-27; issues #10, #12, #13)

• The battery bay (cad/body_battery.py): the two DCB609 packs hang from their rail adapters under the belly, out through two end doors on thumbscrews; 132 mm deep; the robot sits on its skids under a 139 mm leg (H_RANGE 160-375).
• One converter, TRACO TEP 150-4812WI (12 V 150 W, 18-75 V), for the J501 and the 12 V rail; the DDR-120C-24 and DDR-60L-12 are gone. The turret (D66) is a 7.5 A branch of the bus through the stop relay's pole B (the G2RL-2's contact carries 8 A at 30 V, about 6 A breaking at 41 V: B15 measures the inrush and the contact).
• The firmware's 10S windows (merged from body/b8-firmware): warn 35.0, jump gate 34.0, sit 32.0, cut-off 30.0 V under load, 30 A / 80 A current limits, the regen clamp over 41.5 V and the 44.0 V latch; the jump 199 / 150 / 104 mm at 41 / 36 / 34.1 V. The LQR table is regenerated on the D65 model (33.58 kg).
• B9 (the dock) and B17 are overtaken: D64 charged the pack through pads in the tub's floor; tool packs charge on DeWalt's chargers. A parking stand with the user's DeWalt chargers is the obvious successor; not planned until the coordinator asks for it.
• The mass gate fails: P3 33.8 kg against 32.0 (the packs 3.13 kg against the brick's 1.54). The options are in docs/body-sections.md 6; the gate's test is a strict expected failure until one is chosen.
• The B5 fuse board is reworked for 10S (body/b5-fuse-board 760e889, merged): F7 the TRACO 10 A, F8 the turret 7.5 A, F11 the J501 10 A, F12 the other 12 V loads 7.5 A, F9-F10 spare.
• Mechanics first (issue #13): cad/body_joints.py models every screw as a joint and checks it through the CAD; it found and fixed thirteen faults (the commit messages list them); cad/body_sections.py writes docs/body-sections.md.
• The robot's scenarios at 36 V on the D65 model (33.80 kg; tools/simulate.py --v2, exports/body/sim-v2.json): 14 of 15 pass. fall_side fails: with the motors cut and a sideways push, the body lands at 3,408 N against the 3,000 N limit. The 13S tub reached the legs' stops first (25 mm of clearance at the stops); the D65 bay's skids touch the floor under a 139 mm leg, so in a power cut the body lands on the skids at about 1.7 m/s instead of on the stops. The packs then carry that load into their rail adapters. Options (issue #10): crush-pad skids (a 10 mm TPU 85A lattice, modelled with its stiffness in the sim's contact), a shallower bay (packs lying on their side need 122 mm across, more than the 116 mm between the knees: no), or the leg's stops raised to 140 mm (the leg workstream's).

10. D69 in the body (2026-10-28; issues #10, #19, #14, #20)

D69 swaps the BotWheel for a 10 in hoverboard-class hub motor with a 10 x 2.125 pneumatic tyre. It moves the mass gate to 35.0 kg as built, with a 35.5 kg design mass (17.75 kg a leg).

10.1 Where the body sees the wheel

Until the leg's foot change lands (#25), the body draws D69's wheel in the BotWheel's place (cad/body_params.py _fivebar()), using the same solid as sim/wheel_study.py: a 254 x 64 tyre ring, a Ø165 x 65 drum and the Ø17 shaft. The sim (sim/robot_v2.py use_d69_wheel()), the legs' loads (cad/body_legloads.py) and the firmware header (firmware/balance/tools/gen_model.py) put the same wheel into sim.leg_rig's state, at the ledger's 3.53 kg. No leg file is edited.

10.2 The mass gate (issue #10; D71 replaces D69's numbers)

D71 (main e899d28): 37.5 kg design (18.75 kg a leg), 37.0 kg as built. P3 is 36.75 kg, 36.25 kg with the scheduled levers: a 0.75 kg margin, and the test passes. The turret's 2.55 kg head (D66) would use about 0.85 kg of it. What follows is the state under D69.

The P3 ledger is 36.75 kg: 36.25 kg with the scheduled levers, which is 1.25 kg over 35.0. That includes the crush pads' 0.23 kg (10.3); before the pads it was 36.55 / 36.05 kg. The gate is not relaxed: test_the_mass_ledger_under_the_ceiling stays a strict xfail. The levers that could close the gap are listed in docs/body-sections.md 6. The body-only ones are worth 0.8-1.0 kg, which is not enough alone. The rest are:

• the wheel weighing in under its listing;
• the 8.5 in backup wheel;
• lighter packs;
• a ceiling of about 36.1 kg, which needs the leg to re-run its cases at about 18.3 kg a leg.

D69's 35.0 kg came from a ledger taken before D65's packs.

10.3 Crush pads and the lowest leg (issue #19)

The bay's four pads are printed TPU 95A gyroid:

To keep the 40 mm, the lowest controlled leg rises from 160 to 185 mm. H_RANGE in sim/robot_v2.py reads it from cad/body_battery.py PAD; the firmware's GP_H_MIN_M follows, and a test ties the three together. The leg's stops do not move.

Results, with the densified pad modelled as stiff as the 1 ms step allows:

• No 25 mm or 35 mm pad passes all nine at any plateau.
• 45 mm works only in a 600-650 N window, too narrow to build to. The review of #21 found that 600 N reaches its stroke at x1.4, and 650 N peaks at 2,909 N at x0.6.

The robot sits on the pads under a 145 mm leg; they are D69's parking brake. The pads add 0.23 kg against the 30 g skids. Evidence: sim/bay_pad_study.py, exports/body/bay-pad-study.json and bay-pad-depth.json. Still to do:

• Issue #26, before the pads are printed:
  • a coupon crush test and a drop and rebound test (acceptance: a plateau of 500-700 N a pad, at most 24 mm in a 1.6 m/s drop, a rebound at most 50 %, a slanted drop bonded);
  • the retention: a 3 mm lip up the floor flange's outer face, and two M3 into new PEMs in each flange. The model's pads are rigid sideways, and VHB alone may peel at about 1.7 MPa.

10.4 Track, stance and the wheel's connectors

Each wheel connects at the cheek's end, where its lead enters the cable chase beside the RS-04 harnesses. The body carries the sockets and the wheel's lead carries the pins:

Prices are still to read. Mouser timed out for an automated read on 2026-10-28. The BOM gets these lines once their prices are read (D40).

10.5 IP54 (D69: splashes, puddles to 40 mm)

The sealed volume is the shell, the deck and the belly plate. The bay sits outside it: the packs are DeWalt's jobsite-rated tools, and their adapters face up under the belly plate.

10.6 Still open

• The DCB609 and the Terrafirma adapter are envelopes until they are measured (#20). Their rows say so, and a test keeps them labelled.

• The body's strength at D71's 18.75 kg a leg (gb10-02 and gb10-01; exports/body/fea/summary.json, leg-loads.json; 41 V; the side grip at the 10 in wheel's radius and tyre plane). Every case passes:

  Case	SF	Target	Mesh
  S1 landing	2.38	2.0	4 mm
  S1F collapse	1.30	1.2	4 mm
  S1b	2.78	2.0	4 mm
  S2	8.0	2.0	4 mm
  S3	28	2.0	4 mm
  T1 tray	2.61	2.0	4 mm
  D1 front / riser	5.5 / 5.9	2.0	4 mm
  D2 turret fall	2.12	2.0	2 mm

  The 15 scenarios at 37.5 kg (0.75 kg of ballast) pass 15/15 (exports/body/sim-v2-design-mass.json).

• How D2 was fixed (exports/body/fea/D2-convergence.json; tools/fea_convergence.py). With the turret plate bonded over its whole face, its outer corners pried against the deck. Removing the front pockets made that worse (194-202 MPa), so the corners were a model artefact, not the pocket. The plate is now modelled as the bolted joint it is: a 0.5 mm gap, clamped under each riser foot (the M5's preload sandwiches the deck) and at its end nuts.

  • With the Ø36 feet the deck's peak at the foot's rim converged to 107.6 / 104.0 / 100.4 MPa at 4 / 3 / 2 mm (SF 1.92).
  • The riser's feet grew to Ø44, which gives 96.7 / 95.9 / 89.9 MPa, SF 2.15 at 2 mm. D2 now runs at 2 mm.
  • The rigid-pad rims (the posts' corners, the clamp pads) are singular: 105 -> 375 MPa from 4 to 2 mm. Their abuse check is judged at the 4 mm reference mesh only; finer meshes report them, and the bolted joint is the hand check.
  • The wider feet cost 4 of the deck's 31 outlet cells. In the hand model the flow drops from 12.89 to 12.42 CFM, the cooler's inlet rises from 2.49 to 2.59 K over the room (limit 3 K), and the plate goes from 60.3 to 60.5 C.
  • The plate's foot pads are now r 22 (134.5 g, inside the 135 g budget).

• The deck's and the turret plate's SendCutSend prices are to re-read, because their DXFs changed (the cells, the plate's pads). Run tools/sendcutsend_quote.py --only deck turret_plate, then cad/body_bom.py. This session's Playwright could not launch Chrome 151 headless.

• Price the Deutsch connectors, the grommets, the vent plug and the gasket.

• Open B7's louvre filter.