Body plan (milestone M5): the body around two legs

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: docs/requirements.md, docs/decisions.yaml (D28-D41), docs/milestones.yaml, docs/head-payload.md, docs/turret-plan.md, docs/mechanical-interfaces.md, docs/cad-plan.md, sim/robot_v2.py, sim/leg_rig.py, designer/robot_v2.js, catalog/*.json, exports/leg/leg.json, bom.json and rig-results.json, build/sim/v2/model.json, cad/fivebar.py (hip_frame, stack), cad/turret.py (the TM pattern), cad/turret_box.py, research/power-electronics.md, research/lfp-batteries.md, the RS-04 manual (references/vendor/robstride-rs-04), the Mean Well DDR-60 datasheet and the Littelfuse KLKD page (2026-10-27), and the J501 project (~/Code/j501-agx: README, cad/parameters.json; read only). Prices carry the date they were checked; "verify" means the page was not read today. Section 9 lists what is still to verify.

Every "decide" below is decided here, as the user asked (D41: "You don't need to wait on anything for me"). Three of them change the current sim default (designs/default.json); they are written up as proposed decisions D42-D44 in section 7 for the user to record.

0. The plan in one paragraph

Build the body as an aluminium spine box between the two hip frames plus a floor tray, a deck plate and a belly plate, all SendCutSend sheet at an instant price, with printed PA6-CF posts, riser and battery cassette and printed PETG shell panels that carry nothing structural. The two hip frames bolt to the spine's cheeks with the six M5 inserts each already has (cad/fivebar.py hip_frame), 42 mm apart; the gap between the cheeks is the cable chase. The body is 320 × 260 × 140 mm over the hips (the sim's 280 mm interim shell cannot hold the J501 carrier, the battery and the 24 V converter side by side; section 1.9 lays it out). The robot runs on a 13S Li-ion bus (48.1 V nominal, charged to 4.10 V/cell = 53.3 V) with a 100 A-class BMS: the one voltage that keeps the RS-04s at the speed the jump needs (the rig's jump peaks at 16.7 rad/s at 60 N·m), stays under the RS-04's 60 V over-voltage fault with room for regen (so the $444 maxon clamp goes), and feeds the J501 and the turret through the catalogued Mean Well DDR-120C-24. The FLEXVOLT 60 V default (D25) cannot stay: DeWalt's "60V MAX" is the pack's initial no-load voltage, and the RS-04 manual sets its over-voltage fault at 60 V, so a fresh pack faults the leg motors before the first regen pulse. The first physical prototype is the open frame with the battery, compute and power skid and no shell, balancing on a tether; the shell follows.

1. Architecture

1.1 Structure

What the legs give us. Each hip module (cad/fivebar.py hip_frame, exports/leg/leg.json) is a printed PA6-CF "C": a 24 mm web on the leg side holding both RS-04s on their lugs, and two 7 mm flanges (top and bottom, z = ±62 to ±69, x = ±95) that run inward over the motors to the module's centre plane (Y['centre_plane'] = -82.3 mm from the AO face). Six Ø14 bosses on the flange ends, at (x, z) = (0, ±67) and (±83, ±67), carry M5 heat-set inserts (McMaster 94459A775) 11 mm deep, axis along y, facing the centre plane. That face is the only body interface the leg offers; the motors' rear covers sit 2 mm behind it, and the RS-04 harnesses leave those covers radially, i.e. right at the centre plane. The flanges reach only 2 mm above the Ø120 motors, so nothing can bolt down through them.

Decision: a spine box between the frames, 42 mm wide. Two frames cannot bolt to one plate: their bolt patterns coincide (the right frame is the left one mirrored), so the heads collide. The spine is therefore one bent part with two cheeks 36 mm apart; each frame bolts to its own cheek from inside the gap (6 × M5 × 16 button heads, a 4 mm L-key or a stubby driver; the gap is open at both ends and at the top through the floor tray's centre slot). The frames' boss faces then sit at robot y = ±21 mm, the track grows from 373.8 to 415.8 mm (Y_OFF from -82.3 to -103.3), and the gap doubles as the cable chase: the four RS-04 harnesses and the two wheel harnesses (through the frame's Ø18 pass at x = 0, z = 0) enter the gap through windows in the cheeks and go up through the floor tray's slot or down to the belly.

Why sheet and not print for the frame (D33 says every made part is printed; D40 allows SendCutSend): the frame ties two 9 kg legs together through twelve M5 bolts and carries about 10 kg of body with the turret on top; a 320 mm printed box flexes and creeps under the cooler clamp and the leg moments, and the first prototype has to be stiff so the balance loop can be tuned against the simulation. Five laser-cut parts arrive in days at an instant price. The printed body of M5's done_when ("a custom printed body") is the shell around this frame; a printed spine can be tried later on the same hole pattern once the loads are measured.

SendCutSend's rules, read 2026-10-27 (sendcutsend.com/materials, /services/cnc-bending, /services/hardware, /services/countersinking, /faq): 6061-T6 is cut in 0.040-0.750 in but not bent; 5052-H32 is bent in 0.040-0.250 in. For 0.125 in 5052: K-factor 0.44, effective radius 0.125 in at 90°, die width 0.630 in, minimum flange 0.476 in after the bend, maximum bend length 44 in, no custom radii, no angle sharper than 130° or flatter than 5°, cut features at least half the die width from a bend centreline. Flat parts: instant quote up to 30 × 44 in; minimum 0.25 × 0.375 in. Hardware insertion: press-fit nuts S-M3-0, S-M4-0, SS-M5-0, S-M6-0 and M3/M4/M5 standoffs in aluminium 0.040-0.500 in ("starts at $9 per order, many fasteners under $1"); tapping M2-M10 up to 3 × diameter deep; countersinking 90° metric M2-M8 only in ≥ 0.125 in stock. Production: sheet 2-4 business days, bending +1-2, hardware +1-3, countersinking +1-2, anodizing +1-7; no minimum quantity; free US shipping over $39. So the bent parts are 5052, the flat parts 6061, the pallet is 0.125 in for its countersinks, and the whole set ships in about a week.

Load path and the FEA cases (task T6). Per hip: the HM budget of 600 N and 80 N·m (docs/mechanical-interfaces.md) and the rig's real peaks (exports/leg/rig-results.json, 13.1 kg per leg): jump_peak hip joint forces 1.03 kN each (axle 2.38 kN), the undamped power-off collapse 1.43 kN per hip joint, the stop torque 97 N·m; the motors' 2 × 60 N·m reacted by the frame into the cheeks as a couple on the top and bottom boss rows (0.9 kN fore-aft per row); the leg's out-of-plane load (120 N sideways at the wheel). The deck: the turret's 0.05 N·s recoil, a 300 N hand load, 1.5 kg at 120 mm. Targets: SF ≥ 2 on yield (6061-T6 240 MPa, 5052-H32 190 MPa), deck deflection ≤ 0.5 mm under 300 N, the frame's twist between the two hip faces ≤ 0.1° under one leg's 80 N·m. The twelve inserts are the one unmeasured link (M3's coupon test measures an M5 insert's pull-out in printed PA6-CF): the cheeks also bear on the flanges' end faces, so the bolts see mostly shear and the bottom row's tension in a landing (≤ 0.45 kN per bolt at 2.5 kN per leg).

1.2 Battery (D65, replacing D43/D54: rewritten 2026-10-27)

Loads (standing, lights off): 4 × RS-04 at 10 W = 40 W; Orin AGX + J501 ~40 W (15-40 W module modes, 60 W system max); VLP-16 8 W; 2 × D435 7 W; VESCs and wheels balancing ~10 W; MCU, IMU, eyes ~3 W: **110 W**. Playing: plus the turret (D66: its own robot on a bus branch, 40 W average, flywheel bursts to ~140 W) and the SSC1 pods 25.6 W: **190 W**. Peaks: a jump's push draws about 70 A for 0.3 s (the leg rig's HIL at 16.25 kg a leg).

Decision (D65, the user's): two DeWalt DCB609 FLEXVOLT 9.0 Ah packs in 20 V mode, in series: 10S3P, 36 V nominal, 41 V full, 30 V empty, 360 Wh rated, 3.1 kg (DeWalt US: 3.45 lb a pack). Each pack slides onto an off-the-shelf rail adapter (Terrafirma Technology, 10 AWG leads, 30 A stated, no fuse or switch; $64.00 the four on Amazon, B07HYJ628S). 60 V mode is ruled out: a fresh pack sits at the RS-04's 60 V over-voltage fault. One pack (18 V) is under the RS-04's 24 V.

• Where: the battery bay under the belly plate, between the knees (cad/body_battery.py; docs/body-sections.md 1): the packs hang from their adapters on printed plates as from a tool, rails along x, the front one slides out of the front door and the rear one out of the rear, each door on two knurled M3 thumbscrews (tool-free). The bay is a SendCutSend 0.063 in floor and sides on printed rails, TPU skids on VHB under it: 318 × 102 × 132 mm, its bottom at z -224.6. The robot sits on the skids under a 139 mm leg; the controller's lowest height is 160 mm (21 mm off the floor). A pack's 97 mm height is why the bay is 64 mm deeper than the 13S2P brick's.
• Protection is the robot's own (the packs talk to nothing): the firmware's windows (docs/body-firmware.md 5): warn 35.0 V, no jump under 34.0 V resting, sit 32.0 V, a latched cut-off at 30.0 V under load (3.0 V a cell); the bus current limited to 30 A (1 s) and 80 A (0.5 s); regen clamped over 41.5 V; latched off over 44.0 V; no arming over 42.0 V. The MRBF 80 A main fuse and the anti-spark switch stay (D55).
• The jump (the leg rig's HIL with the firmware's current-limited push): 199 mm on a full pair (41 V), 150 mm at 36 V, 104 mm at 34.1 V, refused under 34 V; the 13S pack gave 245/231 mm.
• Charging: the packs come out and charge on DeWalt's chargers; the robot has no charge port. The dock (D18, D64) as planned charged the pack through pads in the tub's floor; with tool packs it is overtaken (docs/body-next-plan.md 9).
• Converters: the DDR-120C-24 starts at 33.6 V, inside the 10S range, so one TRACO TEP 150-4812WI (18-75 V in, isolated, 12 V 12.5 A) feeds the J501 (12-36 V in) and the 12 V rail in place of the DDR-120C-24 and DDR-60L-12.
• Mass: the battery group is 3.27 kg against the brick's 1.54: the P3 ledger is 1.8 kg over the 32.0 kg gate (D57); where the mass comes from is open (docs/body-sections.md 6, issue #10).

1.3 Power distribution

pack (13S, XT90-S) ─ Flipsky anti-spark switch (soft start; 16 mm button on the belly's rear face) ─ MRBF 80 A ─ + bus bar
   ├─ gPV 20 A × 4 ─ each RS-04 (XT30 as shipped; 10 AWG trunk to the chase)
   ├─ gPV 30 A × 2 ─ VESC L, VESC R (XT60, 12 AWG)
   ├─ gPV 5 A ─ DDR-120C-24 → 24 V → J501 DC-IN (Phoenix MSTB 2-pin; polarity checked with a meter first: j501-agx README)
   │                             └→ gPV 5 A (24 V side) → turret DT12 pins 1-4 (≤ 2.5 A; the ramped spin-up is 2.1 A; the stepped one is forbidden on the robot)
   ├─ gPV 3 A ─ DDR-60L-12 → 12 V → lidar interface box (9-18 V, 8 W, 3 A spin-up), SSC1 pods (2 A, PWM-dimmed by a low-side MOSFET), eye controller
   └─ gPV 2 A ─ Mean Well DDR-30L-5 (18-75 V in, isolated) → 5 V → balance MCU, VN-100, CAN boards, INA228, radio-stop receiver
   (gPV = 10 × 38 mm 1000 V DC solar fuses in DIN holders; section "Fusing" below)
 - bus bar ─ 100 A / 50 mV shunt ─ pack
charge port (XT90E panel socket, belly rear) ─ pack side of the switch (through the BMS)

• Voltages: bus 39-53.3 V (RS-04 24-60 V; Flipsky 75100 14-84 V; the anti-spark switch 12-60 V; MRBF 58 V max, so the bus must never exceed 57 V: the VESCs' max-input-voltage parameter at 57 V reduces regen above 55 V, and the 13S charge rule keeps the resting pack at 53.3 V).
• Converters: the DDR-120C-24 is catalogued ($69.20 at Bravo Electro and at Master Electronics, where 235 were in stock on 2026-10-27, though that page also says "no longer being manufactured" with a 19-week factory lead: order two now, one as the spare, and ask Mean Well in T5 whether the DDR-120 is really ending; 33.6-67.2 V in, 120 W, 180 W for 3 s, 102 × 32 × 125 mm, 0.51 kg packed) and keeps the Jetson's ground away from motor ground. Decision: the J501 and the turret share it (as cad/turret_box.py's pod does on the box): J501 ≤ 60 W plus the turret's 40 W average and 2.1 A ramped peaks stay under 120 W, the turret's 3 × 1,000 µF bank rides through, and the stepped spin-up is forbidden on the robot. If the DDR-120 goes away, the SD-100C-24 ($40.70, Bravo 2026-10-27; 36-72 V in, 100 W, 199 × 98 × 38 mm chassis mount) takes its place on the floor tray and T3 reserves its 20 × 10 cm footprint; the DDR-60L-24 ($34.20, Bravo; 18-75 V in, 60 W, 52.5 × 90 × 54.5 mm, 216 g; Mean Well DDR-60 datasheet, read 2026-10-27) is the upgrade if P2's tests show the J501's input dipping under the turret. The 12 V rail is the DDR-60L-12 ($34.20, Bravo; same case, 91 %; the catalogued DDR-30L-12 is too small once the pods' 25.6 W and the lidar's 3 A surge share it). The 5 V rail is the isolated DDR-30L-5 (18-75 V in, 5 V 6 A, 54.5 × 35 × 90 mm; about $27.60 at Bravo, verify today's figure), because every Pololu 60 V-input buck (D42V55F5, D42V110F5, D42V55F12, D42V110F12, D36V50F5/F12) was out of stock or "rationed" with backorders on 2026-10-27. The DDR-60L-12 and the DDR-30L-5 lie on their sides in the belly (52.5 and 35 mm tall); the DDR-120C stands on its 32 mm side on the floor tray at ≤ 50 % load (Mean Well rates derating only for vertical mounting, so all three run well under their ratings; section 1.9).
• Fusing (all DC-rated for the bus): MRBF (58 V DC) 80 A for the main; 10 × 38 mm gPV solar fuses (1000 V DC, 2-32 A; Amazon kits with DIN 10 × 38 holders, about $1-2 per fuse, verify) for the branches: 20 A per RS-04 (its XT30 lead), 30 A per VESC, 5 A for the DDR-120C, 3 A for the DDR-60L-12, 2 A for the 5 V buck, 5 A on the turret's 24 V branch. The branded option is Littelfuse KLKD (600 V AC/DC, 1/10-30 A, fast-acting; littelfuse.com, read 2026-10-27), but at $12-24 each from distributors (findchips/Digi-Key snapshot 2026-10-27) it is the backup, and Littelfuse's SPF PV series (1000 V DC) the middle ground. ATO/MINI blade fuses are 32 V parts and are not allowed on the bus; the turret base keeps its own 10 A blade fuse on its 24 V side.
• Pre-charge: the anti-spark switch's soft start charges the VESCs', RS-04s' and converters' input capacitors; the XT90-S covers hot-plugging the cassette.
• No contactor, no e-stop button (D37): the STOP input (the radio stop, designed separately) tells the MCU to sit the robot down and disable torque; it may also drive the switch's OFF pin after the sit. Flipsky's and Pololu's own notes apply: a MOSFET switch is not a safety device.
• Regen: absorbed by the pack (the charge FET stays closed below 4.25 V/cell); the maxon DSR 70/30 is dropped. The first tethered tests put a scope on the bus during a hard brake from 2 m/s and a full squat to confirm it stays under 57 V. If it does not, the fallback is RobStride's own Discharge Module (Product_Information repo, read 2026-10-27: 15-60 V operating, 80 V for 100 ms, 30 A continuous and 200 A for 100 ms into a 3 Ω/500 W to 10 Ω/200 W resistor; in its "48 V mode" it starts bleeding at 64 V and stops at 59 V, i.e. above the RS-04's 60 V figure, so for this bus its threshold would be set by the resistor-free "48 V mode" only after a bench check; listed at $47.55 by the reseller rcdrone.top, sold out on 2026-10-27, and not found on Seeed's store: verify before relying on it). RobStride also notes that the RS-04's "post-power-off anti-backdrive protection" (its damper register) is what prevents surges when an unpowered motor is back-driven: the leg workstream's D39 bench test should measure it with that setting both ways.

1.4 Controller, IMU, CAN

• Balance MCU: STM32G474 (3 × FDCAN, 170 MHz, FPU; D7's assumption stands). Primary board: ST Nucleo-G474RE ($19.90, in stock at ST's eStore on 2026-10-27; Digi-Key and Mouser would not open) or the catalogued WeAct core; backup: Teensy 4.1 (3 CAN, one FD; $31.50, in stock at SparkFun 2026-10-27). Three transceiver boards: Adafruit CAN Pal 5708 (TJA1051T/3, 3.3 V, CAN-FD to 5 Mbit/s, $3.95, in stock 2026-10-27), which also covers bus C's FD option; the Waveshare SN65HVD230 ($4.99, classic 1 Mbit) is the alternative.
• IMU: the user's VN-100 Rugged (D6) on its 2 × M3 + 2 dowel pattern, flat on the floor tray beside the spine slot (x ≈ -30, y ≈ 0, 100 mm above the COM, under the carrier's standoffs), read at 800 Hz over TTL serial; backup Adafruit ISM330DHCX (raw 6-DoF to 6.7 kHz, $19.95, in stock 2026-10-27). VectorNav lists no price but "ships within 1 business day" (vectornav.com, 2026-10-27).
• Loop: the rig's virtual-model controller (sim/controller_v2.py) at 1 kHz: leg length and lean through the five-bar Jacobian, wheel torque for balance, the TPU stop springs; MIT-mode torque to the RS-04s, current mode to the VESCs. Start-up: switch on → the MCU waits for bus ≥ 39 V → enables motors → the Jetson commands stand-up.
• CAN topology (the J501 has two isolated CAN-FD ports: the Seeed wiki, read 2026-10-27, says "two CAN interfaces that support CAN FD at 5 Mbps", capacitively isolated, a 120 Ω terminator each "installed by default" and toggled by GPIO, both on the J25 20-pin terminal: CAN0_H/L at the end of the top row, CAN1_H/L at the end of the bottom row next to the two GND_ISO pins; the datasheet's "1x CAN" is wrong, and there is no 40-pin header on the J501):
  • Bus A (MCU FDCAN1, 1 Mbit classic): left leg RS-04 × 2 at 1 kHz (~0.5 Mbit/s); 120 Ω at the far motor's lead.
  • Bus B (FDCAN2): right leg RS-04 × 2, the same.
  • Bus C (FDCAN3): VESC L + VESC R at 500 Hz plus the J501 can0 (setpoints 100 Hz, state 200 Hz, 200 ms heartbeat); the J501's own termination is one end, the far VESC the other.
  • J501 can1: the turret's Feather M4 CAN (docs/turret-plan.md D3), the J501 terminating; the turret terminates its own end.
  • Motor-side CAN grounds tie to the J501's GND_ISO.
• Jetson link protocol: velocity, yaw rate, height, lean, "sit", "jump" in; pose, bus V/I, motor temperatures, faults out; heartbeat loss → balance in place for 2 s, then sit. Same for the STOP input.
• Eyes and character: the Waveshare DualEye on an Adafruit QT Py ESP32-S3 (5426, $12.50, in stock 2026-10-27) over USB from the Jetson, off the balance loop (D16).

1.5 Head

As the user set it (2026-10-26) and sim/robot_v2.py computes, moved to the 320 mm deck: the VLP-16 flush with the deck's front edge (axis x = 160 - 51.65 = 108.35, pins fore and aft at x = 63.9 and 152.8, cable exit +y), the turret's pan axis 143 mm behind it (x = -34.65; the sim's 5 mm clearance rule between the lidar and anything that turns) on the 30 mm riser (D41: 13.9° of depression facing forward over the lidar, 25° beyond ±56° of pan, from the sim's riser study); the turret's base reaches to x = -119.3 and the bumper ring to -120.65, 39 mm inside the rear edge at pan 0 (it overhung by 0.7 mm on the 280 deck); the turret hides 84° of lidar azimuth behind the robot at pan 0. The ears are the two SSC1 pods on the side walls (x -10..51, 109 mm behind the front face so the battery door can be in front of them; section 1.9), their tops at least 23 mm below the deck so they stay under the lidar's -15° beam (the sim's ear_top rule for the new lidar position), aimed forward with the bracket's tilt bolt; the DualEye alone on the face at z ≈ 166 (the sim's eyes mount); the front D435 in the front wall at z ≈ 93, the rear D435 in the rear wall; nothing sticks out the front. The VLP-16's 3 m cable is "permanently attached to the sensor" (user manual 63-9243 Rev F §5.4.1, read 2026-10-27; cutting it needs a Velodyne waiver) and ends in the interface box, a Polycase LP-41 enclosure (drawing 50-6001: about 84 × 84 × 32 mm, ~90 g, unconfirmed; RJ45, a 5.5 × 2.5 mm barrel jack, 9-18 V, an internal 3 A blade fuse and 32 V over-voltage trip, 3.0 A at spin-up). Box and coiled cable, about 0.35 kg, sit above the battery cassette (section 1.9); the Ethernet goes to the J501's 1GbE on the right side (a 12-55 V DIN switch such as the SODOLA SL720S-4T1GT, $28.99 on Amazon 2026-10-27, only if a second Ethernet device appears; the TP-Link TL-SG105 is a 5 V unit).

1.6 Cooling

• Orin: 40 W module mode (60 W system max; J501 parameters.json thermal: TTP 80 °C limit, PETG wall limit 60 °C). The Seeed cooler's 8015 fan draws air down through its top grille and exhausts along the fin edges (±X of the carrier, which is the robot's ±x in the layout of section 1.9). The cooler sits under the riser with its fan inside the riser's Ø144 footprint (fan centre x = -60, y = +9 against the pan axis at -34.65): the riser's open sides and a hex field in the deck inside it are the intake, 66 mm above the cooler's top; the exhaust leaves through louvres low in the rear panel and the forward side panels. The j501-agx CFD (tools/thermal.py, cadagent.cfd) is the check (task T7): plate ≤ 70 °C at 40 W in 25 °C air, fan intake within 3 K of the room (Rev 0.8 of the box reached 57 °C with the fan re-breathing 7 K warm air; the open riser should do better).
• VESCs: bolted flat to the belly plate (fins down in the moving air under the body); 100 A parts at ≤ 15 A.
• RS-04s: inside the hip frames between the flanges; the motors' rated 40 N·m is far above the 10-17 N·m standing load; the MCU reads their temperatures over CAN and derates at 70 °C.
• Converters: the DDR-120C on its side at ≤ 50 % load with a side-wall vent above it; the DDR-60L-12 lying in the belly at ≤ 60 % load in the air under the body.
• Ear pods: 12.8 W each in a die-cast housing: a vented printed cowl keeps the fins in free air and fingers off.

1.7 Kid safety

Rules used: ISO 13854 (crushing: ≥ 25 mm for fingers, ≥ 100 mm for hands), ISO 13857 (an opening ≤ 4 mm admits no finger), touch temperatures of 48 °C on metal and 60 °C on plastic (IEC 62115-class limits with our margin), the turret's hard rules (barrel never above horizontal, ≤ 70 fps, the safety line held) and the stops of R37/D39 (sit before any planned power-off; an unplanned cut is the legs' case).

People-near mode: a person within 0.6 m in the lidar or either depth camera (the lidar's rear 84° is blind, so the rear camera counts): the legs hold their length, wheel speed ≤ 0.5 m/s, no jumps, turret pan ≤ 100°/s, lights at 50 %, and the robot sits if someone stays within 0.6 m for 2 s. A hazard review on the P1 frame precedes any session with children.

1.8 Mass and centre of mass

That is 12 % over the mass the leg rig carries (13.1 kg per leg). Budget: body + head ≤ 10.3 kg, robot ≤ 28.5 kg (levers, in order: pocketed deck and belly, a 2.5 mm floor, the 13S1P pack (-0.9 kg), posts merged into the side walls); T3's mass table is the ledger and T4 re-runs the rig at half the real mass from T3 so the leg studies see the true load. If a leg part drops under SF 2 everyday or 1.2 in a rare event, the leg workstream hears about it before anything is printed. Balance (hand estimate from section 1.9's positions): fore-aft moment about the hip axis ≈ +0.1 kg·m over 29 kg → the COM 3 mm ahead of the axle; sideways ≈ -3 mm (the DDR-120C and the 12 V converter on the right, the boards and the lidar box on the left). Acceptance: |x_com| ≤ 10 mm, |y_com| ≤ 5 mm at the 300 mm stance, from the CAD in T3; the COM height stays within 30 mm of today's -21 mm below the hip axis (the riser lifts the turret 30 mm; the belly's 1 kg pulls the other way).

1.9 Layout (robot frame, mm; x forward, y left, z up from the hip axis; the floor tray's top at z = 78.5, the deck's underside at 213.5)

The interim body (280 long) put the FLEXVOLT (87 mm wide) beside the compute. A 160 mm wide cassette, a 176 × 163 mm carrier whose ports need 40 mm of cable room, and a 125 × 102 mm converter do not fit in 280 × 254 together; 320 does.

Ground clearance: the belly box bottom at z ≈ -141; at the leg's 101.5 mm minimum the hip axis is 186 mm up, so 45 mm remain, 74 mm at the 130 mm crouch; the legs' lower links sweep outside |y| = 135 and never reach the belly. The deck height stays TOP = 216 (10 mm of room above the carrier stack was the tightest number: the cooler top at ~150, the deck underside at 213.5); the implementer may raise TOP by up to 10 mm if T3's real envelopes need it.

2. The simplest first body prototype, and the steps to the finished shell

P0, the power skid (bench, no legs). The belly plate with the switch, MRBF, bus bars, fuse holders, shunt and the two VESCs; the floor tray with the converters, the boards, the VN-100 and the compute pallet; one battery cassette. Power up, check every rail, run one RS-04 and one BotWheel from the MCU over CAN, the Jetson link over can0, UVLO and the STOP input. Everything here is reused unchanged.

P1, the open frame (first physical body). The spine with both hip modules (from M3's build), the floor tray, the deck plate on four posts, the belly plate, no shell. The compute pallet and the cassette drop in. The lidar and a mass dummy for the turret on the deck (1.5 kg at 120 mm). The robot stands on a tether (a ceiling line or a gantry with 0.3 m of slack) and runs the controller from sim; tests in order: stand up, hold, squat, drive 1 m/s, push recovery, brake from 2 m/s with the scope on the bus, then the first jump. This is the robot that proves the balance loop, the power system and the frame stiffness; nothing in it waits on a printed shell.

P2, the head. The riser and the real turret on the TM pattern, the DT12 harness, the lidar; the target-practice scenario on the floor with the robot tethered, then untethered in a padded room.

P3, the shell. Side walls with the ears, the door and the port hatches, vents; the front face with the eyes and camera; the rear with the camera, antennas and the radio-stop jack; the belly box with the button and charge socket; skids; the thermal test at 40 W; the hazard review.

P4, the dock (section 2.5).

2.5 The dock (D18), outline for the next session

The robot reverses over a 40 mm tall dock using the rear D435 on an AprilTag and the lidar, then sits. Under the belly plate two brass pads (McMaster brass bar, 25 × 60 × 3 mm, on a PETG insulator) land on two spring contacts in a printed funnel that corrects ±8 mm; the dock's supply is a sealed CC/CV LED driver (Mean Well HLG-320H-54A, output pot set to 53.3 V, current pot at 3 A: $99.00 at Bravo Electro on 2026-10-27, and Mean Well's HLG-320H spec gives the A-type's adjust ranges as 49-58 V and 2.97-5.95 A; the HLG-240H-54A at $73.30 is the smaller choice) with a dock-present signal (a 24 V line through a third contact pair); the robot closes its charge relay (a 10 A automotive relay driven by the MCU, in the charge socket's line) only when it has seen the tag, sits with the contacts closed and reads the dock-present line; the BMS does the rest. The dock's weight-held contacts follow D18; a magnetic pogo pair is the backup if the pads foul. The reload station of D22 (balls poured into the hopper's flap) shares the dock. Everything here is TODO (next session).

Order of ordering (long poles first)

The pack cells, BMS and welder; the SendCutSend set (spine, floor tray, deck, belly, pallet: one cart, 5052 bent parts and 6061 flat parts; the rules were read 2026-10-27); the Mean Well converters and the gPV fuses and holders; the Nucleo and CAN boards; the remaining connectors and McMaster hardware. Printing (posts, riser, cassette, shell) runs in parallel. The hip modules, motors and wheels come from the leg workstream's M3 order. Section 8 is the P1 cart.

3. Highly available components (primary and backup)

Prices: (c) = catalog/.json with its price_checked date; (r) = research/.md 2026-10-25; (d) = read 2026-10-27; "verify" = not read today. Every part is off the shelf, printed, or an instant-price custom part (D40).

Radio stop (D37, designed separately): the body provides a 4-pin rear jack (STOP_A/B dry loop, closed = run, +12 V, GND) and a 30 × 50 mm bay beside the rear antenna bulkheads; an ELRS receiver (RadioMaster RP1 V2, $18.99, available at radiomasterrc.com 2026-10-27; $24.99 with 225 in stock at RaceDayQuads) driving a Pololu RC switch with relay #2804 ($19.95, r) is the simplest non-rated version; no 2.4 GHz wireless e-stop with a dry-contact output was found on Amazon (the nearest is a 310-331 MHz crane remote with an 8 A relay, $93); FORT Robotics is the rated one (no listed price, so outside D40).

4. Interfaces to freeze now

With the leg workstream (cad/fivebar.py, sim/leg_rig.py):

1. Hip frame to body: the six M5 inserts per frame at the centre-plane face, (x, z) = (0, ±67), (±83, ±67), bosses Ø14, inserts 11 mm deep, the six boss faces and the two flange end faces flat and coplanar within ±0.2 mm (a print requirement). The body side is a 3.2 mm 5052 cheek per frame; M5 × 16 button heads, 3 N·m. Loads as section 1.1.
2. Spacing and track: the frames' centre-plane faces at robot y = ±21 mm (SPINE_W = 42 mm); track 415.8 mm; Y_OFF = -103.3 mm. The sim, the designer and the CAD read it from one constant.
3. Harness exit: the RS-04 leads leave the rear covers into the spine gap through cheek windows |x| ≤ 20, |z| ≤ 25; the leg keeps that zone free between the motors' rear covers and the centre plane (the gusset at x = 0, z 28..62 stays). The wheel harness uses the frame's Ø18 pass at (0, 0) and runs inward to the gap. Lead lengths: the RS-04's 310 + 40 mm reaches the belly bus bars (≈ 150 mm) and the chase; its CAN leads get GH1.25 extensions to the MCU shelf (≈ 350 mm).
4. Sweep envelope near the body (robot frame, both sides): upper links |y| 103.3..133.3, reaching z ≤ +31 (the +3° stop plus half a link width) within |x| ≤ 225; front links |y| 135..155; wheel links |y| 157..183; tyres |y| 185.6..230.2; the stop posts stand on the frame's outer face at (±114, 42.9) and (±29.5, -36). So: the body shell above z = +35 may be full width (±130); below that, and the whole belly, stays within |y| < 100 and the hip frames' own envelope; the belly bottom at z ≥ -141 (45 mm of ground clearance at the leg's 101.5 mm minimum).
5. Mass: the body + head budget 10.3 kg; the rig re-run at half of T3's real total is the leg's acceptance for it.

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

6. TM on the head: the 30 mm riser's top face is the TM face at (x, z) = (-34.65, 246) in the robot frame (deck TOP 216 + 30), 4 × M5 inserts on 80 × 80, dowels Ø5 at (52, 0) and (0, -52) in the turret's frame, Ø45 hole; the turret's Ø144 foot and Ø172 bumper ring sit 39 mm inside the deck's rear edge (x = -160); the lidar's axis is 143 mm ahead of the pan axis with 5 mm between them at any pan; the depression table per D41 (13.9° forward, 25° beyond ±56° pan) regenerated by cad/turret_lidar.py for the riser.
7. DT12: the body carries the DT06-12SA plug on a 300 mm pigtail from the deck grommet at (-130, 60), wired per docs/turret-plan.md §2.3: pins 1-2 +24 V from the body's DDR-120C-24 through a 5 A 10 × 38 fuse (≤ 2.5 A; ramped spin-up only), 3-4 GND, 5-6 CAN to the J501's can1 (the J501 terminates), 7-8 the body's STOP loop (closed = run), 11 LINE_SENSE 24 V.
8. Cable routes on the deck: the lidar's cable leaves at +y and drops through the grommet at (108, 95); the turret's harness drops at (-130, 60); nothing crosses the fan intake field inside the riser.

With the J501 (read only, ~/Code/j501-agx cad/parameters.json): the carrier's 4 × M3 mount holes and the module's 92 × 79 tie-down pattern, the cooler 100 × 87 × 34 at PCB (90, 90.5) with ≥ 10 mm above it, 16 mm standoffs for the underside cards (M.2 E Wi-Fi, Mini PCIe, M.2 B), DC-IN 24 V on the Phoenix plug, can0 to the MCU, can1 to the turret, the four RP-SMA bulkheads in the body's rear panel, pin access to the REC/RST buttons through the left hatch, the port windows in the side panels derived from the mated plugs as cad/ports.py does there (never from a receptacle bounding box).

5. The ordered task list for the Opus implementer

Estimates are working days for one agent; "sim time" means waiting on MuJoCo or FEA runs. New CAD goes in cad/body_*.py; exports in exports/body/. Nothing in cad/fivebar.py or cad/turret.py changes without the owning workstream. Keep the design simple and every bought part highly available (D40).

6. Risks

7. Proposed decisions for docs/decisions.yaml (the user records them)

• D42 (proposed): the bus is 13S Li-ion, 48.1 V nominal, charged to 4.10 V/cell; the FLEXVOLT 60 V default and the maxon regen clamp are dropped. Why: the RS-04's 60 V over-voltage fault against DeWalt's 60 V initial voltage; the rig's jump speed; 180 Wh; no pack protection. The user's packs stay useful on the bench and for the turret stand.
• D43 (proposed): the body frame is SendCutSend sheet (a bent 5052 spine and floor tray, flat 6061 deck, belly and pallet) with printed posts, riser, cassette and shell, 320 × 260 × 140 mm over the hips; the hip frames bolt to the spine's cheeks with their six inserts, 42 mm apart (track 415.8 mm).
• D44 (proposed): the J501 and the turret share the body's DDR-120C-24; the turret's CAN is the J501's can1; the stepped flywheel spin-up is forbidden on the robot (the ramped 2.1 A is the design point); branch fuses on the bus are 10 × 38 mm 600 V DC midget fuses.

8. The P1 cart (quantities; the implementer's T8 turns this into exports/body/bom.json with checked prices)

9. Still to verify, and TODO (next session)

• Prices and stock were read on 2026-10-27 for the cells, the JK and Daly BMS, the chargers, the Mean Well units (Bravo Electro and Master Electronics; Digi-Key and Mouser would not open), the Pololu bucks, the Flipsky switch and ESCs, the Blue Sea block and fuses, the Nucleo, Teensy, Feather, CAN Pal, the IMUs, the eyes and their controllers, the cameras, the lights, the RP1 receiver, the BotWheel, SendCutSend's rules and the J501's CAN ports; they are in the text with (d). Still to verify: the DDR-30L-5's price, the gPV fuse kit on Amazon, the exact Daly 13S variant, the copper bar stock number and the 5/16-18 insert for the SSC1 feet, Weidmüller's gPV holder number, and whether Mean Well itself has ended the DDR-120.
• The exact 5/16-18 insert for the SSC1 feet; the VLP-16 interface box's size and mass (measure the user's unit); which FLEXVOLT packs the user has (for the bench), and whether one of his own CAN microcontrollers replaces the Nucleo (D7).
• The dock (section 2.5, T10) and the people-near mode's sensor logic (which runs on the Jetson).