The design notebook.

The robot is designed and strength-checked at 37.5 kg, 18.75 kg a leg. The as-built mass ledger must stay at or under 37.0 kg. This replaces D69's 35.5 kg / 35.0 kg. The body takes its listed savings levers (lighter cheeks, aluminium deck standoffs, the smaller cooler, a 1 m lidar cable, the packs weighed: about 0.8-1.0 kg) as margin, not as a reason to lower the design mass. The packs stay the user's DCB609 9 Ah (D65) and the wheels stay 10 in (D69). The leg's strength cases, the rig predictions and the robot sim re-run at 18.75 kg a leg. The lowest controlled leg length is 185 mm, so the body's 50 mm TPU crush pads (600 N plateau, 30 mm stroke) keep every motors-off fall under 3,000 N (worst 2,724 N).

The ledger with D65's packs and D69's 3.53 kg wheels is 36.25 kg after the scheduled savings (body/program 205b807, issue #10), and the turret as its own robot (D66: Orin NX, its own converter) weighs 2.55 kg against the 1.70 kg head budget: about 37.1 kg in all. D69's 35.0 kg ceiling came from a ledger taken before the packs. The alternatives cost what the user asked for: 6 Ah packs lose a third of the energy, and the 8.5 in wheel saves only 0.46 kg. The RS-04 has the torque (it stands at about 18 N m against 40 rated), so the honest answer is to design for the real mass and keep the levers as margin.

D41 reserved circuits 11-12 of the turret's slip ring for a camera. They now carry power, as a third power pair, because two pairs cannot carry the standalone case (D66: one DCB609 in 20 V mode, 15-20.5 V). A camera on the turret uses the Orin NX on the turret itself (D66), so nothing needs to cross the ring for it. On a single pack, the turret keeps every ring circuit at or under 2 A with a profile driven by the supply voltage: at 15 V the spin-up stretches to 2.4 s and the rate drops to 1 shot/s. On Gyropup's 30-41 V bus, the full rate and speed of D68 apply. The stand's rocker switch becomes a DC switch rated 15 A or more (it carries up to 13.7 A at 15 V).

The turret's power rework (turret/program 3e238a6, issue #2) found 3.0-3.3 A per circuit at 15 V on two pairs. A third pair and a low-voltage profile keep the bought slip ring within its rating without a bigger ring or a boost converter. The camera reservation is moot now that the turret has its own computer.

Each wheel is a 36 V 350 W outer-rotor hub motor of the hoverboard family (15 pole pairs, halls, a hollow D-shaft Ø16 mm with a 13.5 mm flat, 76 mm long, clamped on one side in the wheel link's foot) with a 10 x 2.125 pneumatic tube tyre run at 30-35 psi; the backup is the same motor's 8.5 in version. Bought at a listed price (Robokits RKI-9013 Rs 3,256 plus a pneumatic tyre; e-gizmo 230300029; Alibaba/eBay/Amazon 10 in 36 V 350 W listings confirmed live before ordering); two are bought and measured (mass, Kv, R, L, shaft hardness, run-out) before the foot is cut. The wheel link's foot becomes a 45 mm split D-clamp (4 x M5, inserts; a SendCutSend steel clamp block if the print slips). The Flipsky 75100s drive them in FOC on halls with interpolation, 30 A motor max, 10 A continuous; the balance loop takes pitch from the VN-100 and blends the hall edges with the torque model for wheel speed; an AS5047P on the rotor's outer cover is the encoder upgrade if the bench shows a limit cycle. No mechanical or electromagnetic brake: the balance controller holds while powered (3.8 N m a wheel on 10 deg), the skids hold the sat-down robot, a passive phase-short relay is the add-on if P1 shows a roll-away during the fold. Ingress: the hub sealed (greased shaft bearing, silicone cover seam, sealed lead exit, a vent plug), Deutsch DT/DTP wheel connectors, the belly box at IP54; puddles to 40 mm. The ledger's as-built ceiling becomes 35.0 kg (design mass 35.5 kg, 17.75 kg a leg); the strength cases and the rig re-run at it. Stairs are not a v1 capability; the rolling step is 35-40 mm; a 100 mm kerb by jump is a later firmware milestone.

The user wants air tyres for suspension, puddles and steps. The rig at 18.05 kg a leg keeps every case under the RS-04's 120 N m (109 N m uncapped jump, the same as today) and the knee and axle loads inside the leg's FEA cases (the 150 mm drop's axle spike falls 15 %); the clash sweep is clean; the full-robot sim climbs 35-40 mm kerbs against 25 mm today. The cost is the jump (206 to 167 mm), +2.7-3.6 kg on a 32.27 kg ledger, and a 41 mm taller stance (or a 259 mm standing leg). No robot-grade 8-10 in air-tyre hub motor sells at a listed price (D40); the hoverboard family is single-sided by construction, 36 V like the bus, and proven on VESCs. Study: research/air-tyre-hub-wheels.md, docs/wheel-change-plan.md (issue #18, PR #23).

The launch speed target is 95-100 fps (29-30.5 m/s) for a Hasbro Rival round, the speed of Hasbro's own Rival flywheel blasters (Nemesis 94-101 fps measured, Prometheus 100 fps; sold for 14+). The flywheels run from a fixed regulated rail whose no-load speed is the hardware ceiling: no ball may leave above 105 fps (0.94 J, 2300 J/m^2), under ASTM F963's 2500 J/m^2 for a Rival round (109 fps). The chronograph's hardware trip moves to 105 fps. Unchanged: nothing spins unless ARM is held, the barrel never points above level, the laser warns before firing, and the phase-short brake on release. The docs recommend eye protection, as Hasbro does for Rival. This replaces the 70 fps cap of D29 and the turret plans (the EN 71 0.5 J toy limit, 77 fps, which applies to toys for under-14s).

The user: the turret is for older kids and should fire 'pretty good'. 70 fps was our own margin under the EU toy limit for young children; Rival blasters are a 14+ product at about 100 fps, and the ASTM energy-per-area limit still holds at 105 fps with a 4 fps margin. A fixed rail keeps the ceiling physical (no software setting can exceed it), so the safety argument is the same as before, at Rival's own speed.

The turret is designed and bought for Hasbro's own Nerf Rival rounds (22.2-22.8 mm, 1.83 g average measured; the design ball is 22.7 mm and 1.83 g, as the launch, feed and safety numbers already assume). Off-brand Rival-compatible rounds are fine when they are close enough: 22.2-22.8 mm and 1.6-2.0 g (the feed, bore and speed ceiling are checked across that range). The BOM lists an official Nerf Rival refill and a third-party pack (e.g. Worker Cheetah) side by side, cheapest in-stock first. The speed ceiling is D68's; the barrel-never-above-level rule is unchanged.

The user asked for standard Rival rounds: they are sold everywhere, kids already have them, and they are the round the turret's physics (research/turret-launcher-study.md 2.1, BlasterHub's weighing of Hasbro rounds) was sized on. Third-party rounds vary in size and mass, which moves the speed under the cap.

The turret gets its own compute: a Jetson Orin NX 16 GB module (the user's choice; up to 157 TOPS in Super mode, 10-40 W) on a Seeed reComputer Mini carrier board (88 x 56 x 17 mm, 12-54 V in on an XT30), inside the turret's base. It takes one power input on an XT30, 12-54 V: standing alone it runs from one DCB609 in 20 V mode (15-20.5 V) in a dock in its stand; on Gyropup it plugs into the 10S bus (30-41 V) through its own fused branch (D55). Its servos and flywheels run from a wide-input converter inside the turret. Its interfaces to Gyropup are mechanical (the pan base's bolt circle), power (the XT30 branch), Ethernet to the J501, and the dry safety loop; the turret's ARM line and its physical speed ceiling (D68) stay inside the turret, so it is safe on its own. Kid safety stays a hard rule in both forms.

The user asked for it: two robots, the turret and Gyropup, that combine but are independent assemblies, which also gives redundancy. A turret that carries its own computer, power input and safety can be built, tested and played with on its stand before Gyropup walks, and a fault in one does not take the other down. The reComputer Mini's 12-54 V input covers both one pack in 20 V mode and Gyropup's 10S bus with no converter in front of it.

Two DCB609 packs, each in 20 V mode (5S3P, 9.0 Ah, 180 Wh rated), sit in series in two off-the-shelf 20 V dock adapters: 10S3P, 36 V nominal, about 41 V full, 30 V empty (3.0 V a cell), 360 Wh rated. This replaces D43's 13S2P custom pack. The robot enforces the cut-off and the current limit itself (tool packs leave that to the tool, as D23 found), with the MRBF fuse and the anti-spark switch of D55 kept. Every load is rated across 30-41 V: the RS-04 (24-60 V, regen headroom to its 60 V fault is now 19 V), the Flipsky 75100s (14-84 V), the anti-spark switch (12-84 V), the 12 V rail's DDR-30L-12 (18-75 V); the J501's 24 V converter must change, because the DDR-120C-24 starts at 33.6 V (the body workstream picks a converter rated to 28 V or below). The firmware's and the rig's voltage windows move to the 10S range.

The user chose the pack. 60 V mode is ruled out: DeWalt rates a fresh pack at 60 V, the RS-04's over-voltage fault (D43). 20 V mode in series keeps everything in rating with wide regen headroom, uses the user's tool packs and chargers with no custom pack, and both packs swap in seconds. The cost, from the leg rig at 16.25 kg (rig/run.py, the jump, the rig's 44 V gate relaxed, R_int 0.1 ohm): jump 207 mm full / 180 mm nominal / 162 mm at 33 V, against 245 / 231 mm on the 13S pack; 60 N m is still reached; top speed falls with voltage; 360 Wh rated against 433-480 Wh; about 2.8 kg against 2.0 kg, so the mass ledger must absorb about 0.8 kg.

The body's strength cases (cad/body_studies.py) take each leg's load where it enters the body: the vertical force and the roll moment at the hip frame's mount plane, worked out every step of the rig (sim/leg_rig.py, at half the ledger's robot mass) by d'Alembert on the leg module (each body's mass, centre and acceleration from MuJoCo; the rig's ballast force checks the vertical), plus 10 % of the ground force sideways at the contact patch; not the axle's peak force times the wheel's 187 mm offset. Rare events (the jumps, the 150 mm drop) at SF >= 2 on yield; the undamped power cut (D39) at SF >= 1.2.

The axle's peak is mostly the wheel and the lower links decelerating against the ground: at 16 kg a leg the body sees 205 N m of roll and 559 N at the worst landing, about 0.4-0.5 of the axle force times its offset (2.4 kN x 187 mm). Designing the body for the axle force would have tripled its spine for loads it never sees; the leg's own parts are checked by the leg workstream against their own joint forces (D39).

The pack's + lead goes to a Blue Sea 5191 MRBF block with an 80 A 5181 fuse (58 V DC, 2 kA) within 150 mm of the plug; its - lead through a 150 A / 50 mV shunt (an INA228 reads it) to a Flipsky Anti Spark Switch Smart Enhanced 300A V2.0 (12-84 V, momentary mode) on the low side, then the bus bars (SendCutSend C110 on printed insulators; Victron VBB115040020 as the backup). The MINI 58V branch fuses sit in Bussmann HHM inline holders, rated 32 V by Eaton: each is labelled '58 V fuses only' and a fuse is pulled only with the pack unplugged; a JLCPCB fuse board with Keystone PCB holders replaces them in P3. The aluminium frame is bonded to the - bus. The radio stop (D37) is an ExpressLRS PWM receiver (failsafe: no pulses) driving a Pololu RC relay switch whose contact feeds the coil of a DPDT stop relay (Omron G2RL-2, 12 V) grounded through the MCU's MOSFET: it holds only while the radio and the MCU both say run; pole A is the turret's dry safety loop (TM pins 7-8), pole B its 24 V branch, as the turret's plan asks (docs/turret-next-plan.md 3.2); the MCU senses the coil and sits the robot down when the radio drops.

No compact MINI holder rated 58 V DC is sold; the fuse does the interrupting, so the holder's rating matters only when a fuse is pulled live, which the procedure forbids. Bonding the frame turns a chafed + lead into a blown fuse instead of a live frame. One relay with two poles gives the turret its dry loop and an independent power cut that open together, from the radio or from the MCU.

The pack is 26 Molicel INR-18650-P30B (3.0 Ah, 36 A) in 13S2P, two layers of 13 columns lying across the robot (266 x 76 x 60 mm with the BMS, 1.54 kg, 256 Wh to 4.10 V), welded with 2P nickel strip, in a printed drawer under the belly plate (P3: a bent 5052 box). The BMS is a Daly K-series 13S 100 A Bluetooth + UART (98 x 65 x 14 mm, common port on the negative side), set over Bluetooth (cell over-voltage 4.15 V, under-voltage 3.00 V, charge 6 A, 0-45 C). Nothing is wired to the BMS's data ports, whose ground is B-: a QT Py ESP32-S3 on the MCU board reads it over Bluetooth and puts it on CAN bus C. An XT90-S plugs the pack in; the charge port is a female XT60 to the pack through a 10 A fuse (it charges with the main switch off); the charger is a Luna 48V Advanced set to 90 % and metered at 53.3 V before use (backup: a Grin Satiator set to 53.3 V).

The JK-B2A24S15P the plan named is 162 x 102 x 20 mm and cannot fit any cassette that fits the body; the Daly K is the smart 100 A BMS that does. No ready-made 13S pack with a high-current BMS sells in the US at a listed price (Danenergy ships only in Europe). P30B over P50B: 1.54 kg against 2.17 kg on a robot over its mass budget, enough for about 1.5 h of play, 72 A continuous for the 80 A jump pulses. A wired UART from a negative-side BMS is a path round its switches.

Each of the four stop posts (cad/fivebar.py) is a Ø16 x 40 vented pull-out 52100 dowel (McMaster 98432A386, $5.88; DIN 7979 D, m6, its end tapped M8 for a puller), 23 mm into a blind hole in the hip frame's 24.1 mm web (a light fit dialled on a post coupon, Loctite 638 filling the gap, never a hard press), in a Ø36 boss; the two knee-out posts sit at the motor's ring's edge, so their bosses are blended into the ring and the flange's corner (the most the frame takes and still fits the bed turned: 249.2 of 250 mm) and backed by a block 40 mm deep on the motor side (the hull of the boss and the flange's end face, its face towards the motor at r 66 and rounded into the web, 2.5 mm off the motor's Ø120). Over it, a printed TPU 95A sleeve, Ø32 with an 8 mm wall and a 3 mm cap, from the frame's face to 19 mm above the upper link's inner face, a stretch fit without glue. Each upper link gets a radial stop face either side of its hub (stop_dogs: flat out to 70 mm, its back tapered into the link's edge, its root filleted), so the sleeve is pushed square on, 57.8 mm from the motor's axis (fivebar.stop_arm), and the face does not slide on it. The posts sit 60 mm from the motor's axis and 58 deg past each stop angle (+3 / -97 deg unchanged): both of a motor's posts exactly midway between two lug screws, so the lug pattern, the motors' turns and the dowel holes do not move. The M8 shoulder-screw posts (92981A303) and their M8 heat-set inserts leave the leg. The rig's stop is the new sleeve: k = E A / t at the 57.8 mm arm, 856 N m/rad, 7.4 deg of travel to the pin.

The pre-order review found the old posts' heads inside the link's thickness (the link met steel after 5 mm of TPU, review M6) and the stop's moment tearing the M8 insert out of the printed web (S3). Checking the fix on the real solids found the force about twice what the review took: the link's straight edge met the round post 31 mm from the motor's axis, not 58, so a 60 N m stop pushed 1.9 kN, not 1.0. A steel pin turns in a printed hole as a rigid body, so the web bears p = F/(dL)(4 + 6e/L) at the hole's mouth (6M/(dL^2) + F/(dL) with M about the hole's middle, as the knee eyes' check takes it; with M at the face, as the review took it, it reads 3F/(dL) low): the proposed Ø12 pin 22 mm deep came out at SF 0.6 everyday against the link's edge, 1.2 with a radial face (target 2). With the face (F = T / 57.8 mm) and the contact low on the pin (19 mm of sleeve, e = 10.5 mm), a Ø16 pin 23 mm deep bears 19 MPa at the 60 N m cap (SF 2.11 on the humid 40 MPa) and 38.7 MPa in the rig's undamped power cut at D71's 18.75 kg a leg (122 N m, SF 1.03 against the fault case's 1.0; 109 N m and 1.15 at 14.75 kg). A Ø20 pin with an 8 mm sleeve does not fit between the lug screws, and the knee-in posts cannot move down (the front link's apex eye passes 3.8 mm under their sleeves at the squeezed crouch). The sleeve keeps the rig's stop stiffness where it was (856 against 816 N m/rad): a stiffer stop raised the power cut's knee force (1,360 N m/rad: 2.1 kN), which the knees' 6201s cannot take. The frame round a knee-out post, with the pin bearing as a contact: SF 0.76 at the 60 N m cap with a round Ø28 boss (its notches); the ear's lobe alone 1.64 (judged 4 mm beyond the hole, where the pin's own bearing is checked: the stress in that band is in the print's layers, 13 MPa in them against 3 across them). The backing also buries the top flange's end, whose root the lobe loaded in the frame's own landing case (SF 1.33 before the stops were redone, 1.13 with the lobe alone, 1.22 with the rounded backing; a backing cut close round the motor thinned to a knife edge there, 0.43). The stop faces' backs taper into the link's edge with a filleted root: ended square, they cost the link's landing case 15 %; tapered, it is the same with and without them (1.37 and 1.36 at a 3 mm mesh). At 18.75 kg a leg (D71, 2026-10-28): the frame round a knee-out post 2.38 at the cap (E3o) and 1.17 in the power cut (R3o); round a knee-in post 2.70 (E3i); the upper link's stop face 7.36 (E3) and 3.61 (R3). Adds 229 g per leg (the pins 91, the stop faces 79, the frame's ears 38, the sleeves 21; the leg 9.71 kg with the 2026-10-28 fillets) and $6.90.

The hardware arm chain keeps D7, D9 and D10 and is built as docs/turret-wiring.md section 3 draws it, part by part (exports/turret/wiring/parts.md). (1) The logic that makes ARM (the two one-shots, the trip latch, the AND gate) runs on the Feather M4 CAN's 3.3 V, so nothing over 3.3 V reaches the Feather; the servo bus returns to RX through a divider, and the laser's FET gate is driven at 5 V by the bus buffer's spare channel. (2) The base's NE555 runs on the pendant loop's 12 V (no 78L05) and drives the 4N35 at 2 mA through 2 x 2.2 k 1 W, with a diode that keeps a +24 V fault out of it. (3) The six brake FETs short each motor phase to ground (not to a floating star), driven on the cartridge board by a 74HCT14 from ARM through a 6.8 k / 1 uF delay (5-14 ms) with a diode for an immediate release. (4) The Feather's own 120 R is the turret's only CAN terminator. After an independent review against the datasheets: (5) the trip latch is set-only (its D tied high, clocked by gate 2 while the window is open), so no later edge can clear a trip; (6) the Feather reads ARM and the pulse train through 100 k and cannot drive them; (7) the one-shots are the CD74HC423E, which a rising reset never triggers, and an MCP120 makes the power-up reset; (8) each brake FET has 100 nF from gate to source so the ESC's phase edges cannot turn it on; (9) AM32's complementary PWM is off and the firmware ramps between set speeds, so no wheel energy goes back into the buck-fed 12 V rail, with TVS diodes on both 12 V rails and Pololu's Schottky across the 2815; (10) the pendant plugs into a keyed mini-XLR socket that no headphone plug fits, through a 0.1 A polyfuse.

The plan's version would have failed on the bench: its 5 V logic put 5 V on the Feather's pins (not 5 V tolerant) and its gates' 3.3 V pull-ups were under a 5 V 74HC input's threshold; a 5 V 555 drives the opto's LED at about 0.5 mA, too little for the 4N35 to pull its input low reliably; a floating star takes the FETs' gate drive from wherever the ESC's phase-sense dividers hold it, so the brake could stay off; and the Feather already terminates CAN. Shorting the phases to ground is what AM32's own brake does, so the two never fight. The review found that the plan's latch (D = the window) could be cleared by the next slower ball, that the Feather could raise ARM through its own pins, that a '123 one-shot fires on a rising reset, that the ESC's dV/dt would lift the brake FETs' gates through 1 k, that complementary PWM regenerates into a rail a buck cannot absorb, and that any audio plug would have closed the pendant loop.

The two output adapters per leg (between each RS-04's output flange and its upper link) are CNC 6061-T6 from JLCCNC (backup Xometry), uploaded as exports/leg/cnc/output_adapter.step with the callouts in exports/leg/cnc/README.md (Ø60 j6 journal for the 61812, H7 dowel holes, M5 threads, ISO 2768-m). The standard face (cad/interfaces.json AO-L) changes with them: 8 x M5 tapped at 0, 30, 90, 120, 180, 210, 270, 300 deg and 2 x Ø5 H7 dowels at 60 and 150 deg on PCD 50, each 15 deg from every RS-04 output screw; the motor's 9 screws are M5 x 25 socket heads (McMaster 91290A252) in Ø10 pockets with their heads 0.6 mm below the face, 6 mm into the motor's 6.5 mm threads. The upper links' hubs take the new pattern. made: 'print' in the standard still gives a PA6-CF adapter (inserts in the face) for a fit check, not for loads. 154 g each against about 70 g printed: +0.33 kg on the robot.

Held where it really sits on the motor (its contact face and the motor's pins, with the screws in clearance holes), the printed adapter came out at SF 0.86 in the landing case across its layers (1.04 if its 61812 is counted as a stiff radial support), against 1.2, with the 13.1 kg loads (about a tenth worse at the 14.75 kg design mass); the earlier 1.15 came from clamping the screw holes along their whole length. Checking the CAD for the fix found three faults in the printed design: both face dowels sat inside bolt holes (so there were none, though the hub joint check counted on them), six face inserts broke into the motor screws' counterbores, and the motor screws' heads stood 0.2 mm (1.2 mm with washers) into the upper link. Machined, the adapter is at SF 7.6 (landing), 15.8 (a motor at 120 N m against a stalled leg) and 22 (everyday); the joint screws can take full preload; and the bearing journal and dowel holes are held to a fit a print cannot hold. D40 allows instant-price CNC; one part type, four per robot.

One Mean Well DDR-120C-24 (isolated, 24 V 5 A) feeds the J501 and the turret, each through its own 5 A fuse; the turret's CAN is the J501's can1 and its SAFETY loop is a radio-stop relay's contact; the turret's stepped flywheel start is forbidden on the robot (the ramped start is the design point). A DDR-60L-12 makes 12 V (lidar box, radio stop, the P3 lights and eyes) and a Pololu D36V28F5 5 V from it. Every branch on the 48 V bus has a Littelfuse MINI 58V blade fuse (0997 series, 1 kA at 58 V) instead of the plan's 10 x 38 mm midget fuses, which stand 83 mm off their rail and need 193 mm of it. This is the plan's proposed D44, with the fuses changed.

120 W covers the J501 (60 W) and the turret (60 W peak, 40 W playing) with the flywheels ramped; one converter and one fuse type keep the skid on one plate. The MINI 58V is the only compact blade fuse found rated 58 V DC in every size needed (2-30 A).

The spine is two identical C-channel cheeks (5052-H32 3/16 in, 184 x 181 mm webs, 30 mm flanges outwards, eight PEM SS-M5-0 each), 31 mm apart on eight M4 x 31 aluminium hex standoffs (McMaster 95947A065) that stand in the hip frames' bolt rows (z +-67.5, x +-10 and +-72) beside the frames' six M5 inserts (fivebar.frame_mounts(): x 0, +-83, z +-70). Each hip frame bolts to its cheek with six M5 x 12 ISO 7380 button heads from inside the gap. The floor tray (5052 1/8 in, 15 mm lips) lies on the upper flanges at z 90.5, the belly plate (6061-T6 0.100 in, 116 wide: the knee eyes keep 25 mm) under the lower ones; the deck (5052 0.100 in) on four printed PA6-CF posts at z 216 (TOP, as the sim). The spine is 40.5 mm wide between the frames' faces: the track is 414.3 mm (the plan's 415.8). The battery lives in a drawer under the belly plate, the power skid on its own plate on printed posts over the tray's front half, the compute on its pallet over the rear half, the lidar 5 mm behind the plan's spot (x 103: its front pin 8 mm from the deck's bend). This replaces the plan's proposed D43 (a 36 mm bent U spine of 1/8 in sheet: SendCutSend cannot bend a channel whose base is under twice its flanges, and 1/8 in cheeks fail the landing case: SF 0.6).

cad/body_studies.py at the rig's loads on the body for the 32 kg robot (cad/body_legloads.py: 205 N m roll and 559 N at a 150 mm drop, 275 N m and 964 N in the undamped power cut, taken at the hip frame's mount plane): S1 landing SF 4.4 away from the standoffs' bonded ends (1.5 at those ends, a modelling singularity still under yield), S1F fault 2.9 (1.7), S1b one leg against a clamped cheek 2.8 (1.7), the 80 N m twist 0.013 deg, the pitch couple SF 27; the tray with what stands on it at the landing's 8.8 g SF 2.5 (1/8 in: 0.100 in was 1.55); the deck under a 300 N hand SF 2.5, 0.33-0.38 mm; a 30 g tip-over of the 1.70 kg head SF 1.3 against the abuse target 1.2 (exports/body/fea/summary.json). With 1/8 in cheeks no standoff layout passed; the standoffs carry each leg's roll moment straight across to the other frame. One cheek part serves both sides, as one hip frame serves both legs.

The robot runs from a 13S Li-ion pack on a 48.1 V bus, charged only to 4.10 V a cell (53.3 V) by a charger set to that voltage, with the BMS's cell over-voltage at 4.15 V as the backstop; empty at 3.0 V a cell (39 V). The FLEXVOLT 60 V default and the maxon regen clamp are dropped from the robot (the user's FLEXVOLT packs stay on the bench and the turret stand). Every load is rated across 39-53.3 V: the RS-04 (15-60 V), the Flipsky 75100 (14-84 V), the anti-spark switch (12-84 V), the Mean Well DDR converters (33.6-67.2 V and 18-75 V), the MINI 58V and MRBF fuses (58 V); regeneration has room to 57 V. This is docs/body-plan.md's proposed D42, renumbered.

The RS-04 faults on over-voltage at 60 V, which a FLEXVOLT pack's 60 V initial voltage reaches; the 13S bus keeps 7 V of regen headroom under that, gives the rig's jump speed, and puts a BMS's protections on the pack. 4.10 V a cell trades about 9 % of the capacity for cycle life and margin.

v0 is the turret as drawn on main (the slip ring and continuous pan, the SendCutSend aluminium clamp, the pulse-train safety line, the two chronograph gates and the hardware trip, the phase-short brake, the hard tilt stops, the ball + 2.0 bore) on its printed bench stand, with three swaps. (1) Two SunnySky X2216 880KV V3 short-shaft outrunners replace the delisted DJI Snail 2305s, on the same fixed 12 V flywheel rail: Ø34 tyres first (worst-corner ball 64.1 fps), Ø36 only if every ball of the bench ceiling test on Ø34 stays under 60 fps; each hub clamps on the motor's own bolt-on prop adapter (M5 stud, Ø5 pilot, Ø17 flange) under the bell. (2) A third Feetech STS3215 on the servo bus steps the dial, replacing the DJI M2006 + C610 (out of stock; the C610 is $250-320 elsewhere); it takes the switched 12 V rail through a 3 A polyfuse, so the safety line still cuts it, and neither 24 V nor a CAN node reaches the cartridge. (3) The red Quarton VLM-650-03 LPT (Class 2, under 1 mW, Ø7 x 21) replaces the green VLM-520-73 LPT (0 stock, 4 weeks). Not fitted on the bench: the TPU bumper, the fill flap and its spring, the lights and buzzer, the hopper-full beam. The stand's rocker is the DC-rated Switch Components RA1-1A-DC-2-B-1 (Ø20.2 round snap-in). The +-170 deg cable loop stays the documented fallback if the ring fails its bench tests (docs/turret-prototype-plan.md 2.4). D41 stands for the robot.

The DJI parts and the green laser cannot be bought today. The substitutes are in stock at US sellers, sit inside the cage's KV band on the fixed 12 V rail (790-910 rpm/V), and reuse what the turret already has (one servo type on one bus). Everything else is built, swept and documented, so the bench turret is the robot's turret.

docs/turret-plan.md section 1, as its planner recommended, except where D40 overrides it. D1: continuous pan through the slip ring (the plus or minus 170 deg cable loop stays a CAD option). D2: the SparkFun ROB-13065, no quotes (D40). D3: an Adafruit Feather M4 CAN Express on the turret. D4: the pan servo's bus crosses the ring. D5: one fixed 12 V regulator for the flywheels (Pololu D24V150F12 class), which is the speed ceiling. D6: two ring circuits reserved for a camera. D7: the safety line cuts the flywheels and the dial in hardware; pan and tilt stay live, capped at 100 deg/s while disarmed. D8: dropped (no quotes). D9: a phase-short brake on release plus the ESCs' brake-on-stop. D10: a hardware chronograph trip at 66 fps. Also: the turret stands on a 30 mm printed riser on the top deck (14 deg of depression facing forward over the lidar, the full 25 deg outside plus or minus 56 deg of pan), and the pan bearing is a chrome-steel 6810-2RS from an off-the-shelf seller, not McMaster's $195 stainless one.

The user asked not to wait on them for decisions.

Each part of Gyropup is one of three kinds. (1) Off the shelf, with a listed price and a buy link: McMaster-Carr first for hardware and bearings; Digi-Key, Mouser, SparkFun, Adafruit and Pololu for electronics; the makers' own stores for motors, servos and the wheel. (2) 3D printed: PA6-CF for structure, PETG, TPU. (3) Custom parts ordered online from our drawings at an instant price: SendCutSend (laser-cut, bent or machined sheet and plate), Xometry or JLCCNC (CNC), JLCPCB or PCBWay (boards). Nothing needs a request for quote: a part sold only on quote is replaced by one of these. So the turret's slip ring is the SparkFun ROB-13065 (Senring, 12 circuits x 2 A, $27.95, in stock) on the bench and on the robot; the 200 ms flywheel soft start keeps every circuit under 2 A (docs/turret-physics.md V3: 2.2 A peak with the ramp), with power circuits paralleled.

The user wants a robot anyone can order and build: off-the-shelf parts, printed parts and SendCutSend-style parts; vendor quotes (the plan's robot-grade slip rings from Moog, Servotecnica, Senring and Moflon) are not needed.

The rig (sim/leg_rig.py) carries the real 13.1 kg (half the 26.2 kg v2 robot; it carried 8.5 kg before: a lost ballast mass) and records the gearbox torque and every joint's force at every step; the strength cases (cad/leg_studies.py) take their loads from those, not from statics on the axle's peak force, which overstates a landing's knee load about twice. Changes from them: each upper link's knee eye has two 12 x 18 x 20 steel sleeves (40 mm) in a boss 20 mm proud of its inner face, with a 90 mm knee shoulder screw; the 61812 hip bearings are pressed in from the motor side onto a lip (no retainer screws); fillets where the knee bosses and the eye boss meet their links; each chain is checked at its measured share of the out-of-plane load plus 0.15. The leg has no spring for now. The one case it does not pass is an undamped power cut while standing: the leg folds onto its stops with 3.0 kN on each knee; with the RS-04's power-off damping it is 0.3 kN. A bench test of that damping (back-drive an unpowered RS-04, measure torque against speed) comes before anything else is built for it. If it is too weak: the knee-to-knee spring add-on (research/leg-suspension.md section 8), softer longer-travel stop bumpers, and a firmware sit-down before any planned power-off. Re-checked 2026-10-28 at D71's 37.5 kg (18.75 kg a leg), the 10S 36 V bus (D65) and D69's wheel. The rig takes the 10 in wheel's mass (4.0 kg; 3.53 kg listed), radius (127 mm) and air tyre (10 x 2.125 at 30-35 psi, about 60 N/mm: solref 0.0141 / 0.5, research/air-tyre-hub-wheels.md 3.2). The CAD keeps the BotWheel's interface for geometry, because W2's 45 mm split D-clamp foot is not designed yet: its boss in the wheel link's foot, the wheel's mid-plane 37 mm from the link's middle. W3 re-runs the studies with the clamp, the shaft and the measured offset (issue #25). The rig: drop_150 axle 1,605 N; the capped jump's landing knees 574 N dynamic (773 N at 16.25 kg on the solid tyre); the jump 153 mm; the gearbox at most 110 N m; the undamped power cut 2,048 N on a knee and 122 N m on a stop post. The air tyre makes the landings about a quarter kinder. The bigger wheel makes the everyday turn (E1) the sizing case: its sideways grip acts 127 mm below the axle, not 85.6. Changes: the hip frame's gussets run out tangent to the web (their sharp tip was real, not mesh noise: E1 1.43 to 1.77 at 17.75 kg); the front link's knee boss fillet on its outer face goes from 6 to 16 mm, plus a 6 mm fillet where the apex eye leaves its inner face (E1 1.62 to 2.16 at 17.75 kg); the hub dowels are Ø5 x 20 (McMaster 91595A356), 12 mm in the link. Strength at 18.75 kg (exports/leg/fea/summary.json, 4 mm mesh). Whole leg: R1 2.40, E1 2.85, R3 1.88. Upper link: R1 1.83, E1 2.13, R3K 1.08, R2 1.64, E2 3.28, E3 7.36. Front link: R1 1.69, E1 2.00. Wheel link: R1 2.74, R1w 1.61, E1 3.15. Hip frame: R1 1.55, E3o 2.38, E3i 2.70, R3o 1.17. Output adapter: 11.2 or more. Joints: hub dowels 2.95, knee pins 1.31 or more, eye pressure 1.06 or more, stop pin 2.11 everyday and 1.03 in the power cut. Two cases miss their targets. (1) The hip frame's E1 is 1.65 against 2.0, at the bottom flange's square end, on its root on the motor side, across the print's layers. It was 1.98 at 32.5 kg on the BotWheel and 1.75 at 17.75 kg. Five changes there each made it worse or moved the failure, so none was kept (cad/fivebar.py GUSSET_RUNOUT): a trimmed root fillet, a variable-radius root fillet, a web lobe, a backing block and a thicker flange. A snipe gave 1.81 but took R1 to 1.35. The fix, before the robot runs: proof-load the first printed frame on the rig to twice the E1 load at the hip (issue #29). If it shows damage, the frame's flanges move to a bent 5052 part. (2) In the undamped power cut, the rear knee's 6201 is at s0 0.94 against 1.0. The rig folds the leg onto its stops with all 18.75 kg on it. On the robot, the body's skids and D71's crush pads land first, at a leg of 119 mm or more against the stops' 88-101 mm. Over C0 by 6 % is raceway brinelling, not breakage. The fix: the rig's power_off test runs with its bottom bumper set or with the RS-04's measured damping (power_off_damped: 364 N), and the knees are inspected after any undamped cut, as above.

The gearboxes peak at 64-106 N·m in every rig event (under the RS-04's 120). A spring between the upper links, the survey's first answer, cannot be built: in 3D its line passes through the hips below a 250 mm leg. The knee-to-knee spring clears the CAD and holds the unpowered robot up, but costs brackets, a guide and a boot (0.45-0.6 kg per leg), a steady 8 N·m on each knee bolt and 7 mm to the tyre at full crouch; for landings the tyre and firmware matter more than any spring.

Each leg is a symmetric five-bar (cad/fivebar.py): two RS-04s 150 mm apart in one printed hip frame (each on a 61812 support bearing), two 150 mm upper links, two 250 mm lower links meeting at the wheel axle; the wheel link carries the BotWheel on a hollow hub, the front lower link rides on that hub on a 61812, the knees are the standard cartridge (two 6001s on a 12 mm shoulder screw). TPU-sleeved stop posts limit the upper links (+3 deg / -97 deg). Leg 130-380 mm, at least 467 N of push anywhere in that range at the 60 N·m cap, 16.7 N·m per motor to hold 130 N, every joint at least 33 deg from a dead point, every link printable on a 256 mm bed. The track drops from 692 mm to about 360 mm. The first build is one leg on a vertical-rail test rig (sim/leg_rig.py), then the turret, then the body around two legs.

The user asked to focus on one great leg (and the turret) before the rest. The coaxial hip-and-knee stack of D31-D33 put two 56 mm RS-04s side by side on each hip and made the robot 740 mm wide. A sweep of 330 printable five-bar geometries (tools/leg_synthesis.py: 36 feasible) found one with the motors inside the body, the load shared by both motors and links that mostly carry tension and compression, which suits printed parts. The survey (research/wheel-legged-survey-2026-10-26.md) recommended Diablo's coaxial layout instead; with RS-04s that layout is about 630 mm wide, and the five-bar is the usual leg on RoboMaster balance robots, the closest match to Gyropup's job. The price is no 360-degree hip, so the legs cannot swing over to self-right.

The robot carries no e-stop button and no contactor. A radio stop is designed separately. A power cut while standing is the knees' governing case (D39): the radio stop should sit the robot down before power goes.

The user will use a radio and work the stop out separately.

Gyropup is now a single design. Removed from this repo: the choose-your-own-adventure designer pages, the results/comparison page, the design variants, the overnight optimiser loop (iterate, trial evaluation, the 4090 mirror), the catalogue's 305 alternative parts and their product images, and the vendor files of motors we do not use. Everything is kept at the git tag explorer-snapshot-2026-10-26. The generic part of it (catalogue schema, design choices, pluggable evaluation, comparisons) is being turned into a modular comparison framework in cad-agent, by a PR into dev. The old serial-leg robot (cad/build.py, cad/leg.py, sim/model.py and the 3D game page that draws it) stays only until leg v2 is on the full robot, then goes too.

The user: the any-parts concept is cool but not realistic; it belongs in cad-agent as a framework for comparisons, and this repo should build one great robot.

tools/tts_server.py is Gyropup's Piper sidecar, speaking the protocol talk.js already used (GET /health, GET /tts): en_US-hfc_female-medium, lifted 2 semitones by resampling after Piper speaks that much slower, at 0.95 pace, each line cached on disk. /health tells the page to play the audio as it is and names the voice, which the talk dock shows. Voice and settings are flags or GYROPUP_TTS_* variables. Runners-up: en_US-libritts_r-medium speaker 8329, en_US-lessac-medium, en_GB-cori-medium, en_US-ljspeech-medium. The browser fallback is gentler (pitch 1.3, natural voices first). Moxie's Whisper and the brain chain of D27 are unchanged.

The user found the voice odd: with no Piper running, the page fell back to the browser's speech at pitch 1.7. Eighteen single-speaker English Piper voices said the same eight Gyropup lines, 1,047 speakers of the multi-speaker voices were screened on two of them, and each was measured for pitch, intonation, pace and cleanliness; hfc_female, from a studio corpus, is the liveliest with the cleanest pauses. Its data is CC BY-NC-SA 4.0: fine for this private, non-commercial robot, but a commercial product would switch (LibriTTS-R is CC BY 4.0; cori and ljspeech are public domain).

The physics fires Rival-size foam balls (22.7 mm, 1.83 g, Cd 0.5) at 70 fps (21.3 m/s, the kid-safety cap for balls) from a 30-ball hopper: the ammo of research/turret-launcher-study.md section 8 and cad/turret.py. The matched flywheels give them no spin, so they fly on drag alone. On the 3D page they turn, bounce and roll, the turret panel counts the hopper, a shot sounds like a dial step and a flywheel pomp, and Gyropup talks about balls. This replaces the dart parts of D19 and D26; the knock-down targets and the laser rules stand.

The user saw darts in the simulator while the CAD turret shoots balls. A ball leaves slower than the 80 fps dart, so it drops about a quarter more over 4 m, but with 1.6 times the momentum it knocks plates over sooner: the targets scenario downed all five plates with 9 hits from 11 shots (darts: 18 hits from 18 shots, four plates).

All custom parts are printed: structure in PA6-CF on a hardened nozzle (100 % solid where the CAD has walls, so the FEA sees what is printed), the rest in PETG, the lidar guard in TPU. Bought hardware only: bearings, steel pins and shoulder bolts, rod ends on M8 threaded rod, screws, heat-set inserts. Strength uses humid, print-knocked-down PA6-CF allowables (x0.6 x0.8 on Bambu's dry TDS; along or across the layers by print orientation), targets SF 2 everyday and 1.2 in falls, leg flex at the wheel <= 4 mm. Four RobStride RS-04 (hip and knee on each leg; RS-03 fits the same faces), capped at 60 N·m in firmware so the printed joints never see more; the output joint carries it by screw clamp plus the motor's pins (SF 3.2). Hip ab/adduction is not added: two wheels 600 mm apart need no sideways balance and differential leg height levels the body (as DDT Diablo and Ascento do).

The user wants to print it and has PA6-CF on a hardened-nozzle printer; RS-04s ship sooner than RS-03s (Seeed: 11 on hand and 80 due Nov 14, against none until Nov 20). The aluminium parts of D31/D32 were sized for CNC and SendCutSend, which the user does not want.

Applied the review's blockers before optimising: the knee crank and shin lever point 75° from the perpendicular towards the foot (the rod's transmission stays >= 0.5 over the knee range; it was at its dead point), the knee motor is modelled as it drives (a tendon over hip + knee), the body has collision shapes, and falls, a kerb drop and a landing are scenarios. The yoke is a bent hat channel (496 g, was 961 g) and the motor plates are pocketed (344 g, was 418 g); the robot is 21.4 kg. tools/iterate.py (NSGA-II: mass, cost, motor load, width) runs every candidate end to end on the 4090 overnight; the open questions in docs/cad-plan.md section 9 take its defaults until the user answers.

The user asked to iterate on the CAD all night with the best, most available parts (in stock and rated for 60 V: RobStride RS-03/RS-04, catalog/availability.json), on the RTX 4090 (its LLM services stopped with the user's approval). The review (research/mech-review-2026-10-25.md) showed that optimising the old model would have optimised the wrong machine.

Legs are built from motor-specific adapters onto standard faces (GMI, docs/mechanical-interfaces.md): an output adapter (AO-L) with a 61812 support bearing in a motor plate (AC-L), standard faces for links, a yoke box closed by three posts, a 3 mm 6061 chassis as the spine. The thigh and shin are 6061 boxes (40 x 30, 3 mm walls) with the wheel's cables inside. Any of the RS-03, RS-04, CubeMars AK70-10 or Damiao J10010 fits the same leg with its own two adapters.

The user asked for real motor attachment points, real cable routing and stiffness studies, with modular interfaces so motors and leg designs can change. The studies showed printed PA6-CF links and a flat 16 mm thigh far too flexible for the hard-turn case (out-of-plane 50 N m at the hip, 120 N sideways at the knee: safety factor 0.11, 38 mm deflection), and RobStride outputs ride on thin 45 x 58 mm ball bearings, so a separate support bearing takes the leg's bending.

The tail is removed from Gyropup: from the design, the catalog, the physics model and controller, the scenarios, the 3D views and the CAD. Gyropup is back to two balancing wheel-legs. An emergency stop is a controlled crouch (category 1) before the motors are cut, with the knees landing on their mechanical stops; a skid or bumper under the chassis protects the lidar if the robot does tip.

The user found the tail ridiculous. Its benefits in simulation (catching hard backward shoves, a tripod to sit on) did not outweigh how it looks and the extra 1.2 kg.

Beni (Mondo Robotics, 1.75 kg) joins the reference robots for character and tricks (self-righting, hops, follow-me filming, swappable battery). Gyropup stays a ~19 kg outdoor robot and aims higher on durability and capability (see research/reference-robots.md)

The user pointed to Beni as consumer inspiration but wants Gyropup way heavier, more durable and more capable.

Gyropup keeps its two balancing wheel-legs and gains a tail: a RobStride RS-06 lifts it (36 N·m peak, 621 g), a RobStride RS-00 wags it (14 N·m, 310 g), a 20 mm carbon tube carries a TPU skid shoe 390 mm from the pivot, and a spring-applied ratchet (ASSUMED part) locks it against being pushed up whenever the solenoid is unpowered. The tail is not driven. It rides up while balancing; the robot sits on it (2.4 N·m on the motor, parked on the ratchet), it catches a backward fall, and an e-stop is a controlled stop: sit, lock, cut the power.

A third wheel-leg would add about 4.3 kg and six-hundred-odd dollars, make turning scrub, and still balance when lifting a leg; the tail gives the robot most of the tripod's benefits for 1.2 kg. In simulation the tail lets the robot survive a backward shove of 2.0 m/s per kg (vs 1.2 without), sits and stands up again in about 5 s, and stops safely with the lidar 80 mm off the ground. RS-02 and RS-00 pitch motors also sit and catch moderate shoves but drop the hardest catch, so the RS-06 stays. The ratchet sees at most 43 N·m.

designer/character/talk.js gives speech bubbles over the head, spoken lines, listening and a small brain. It uses Moxie's local Piper TTS (:8081) and Whisper STT (:8082) when they run, else the browser's speech synthesis and Web Speech recognition; an OpenAI-compatible endpoint or local Ollama when reachable, else a scripted kid-safe brain. Sound effects (flywheels, shots, hits, knock-downs, laser, servos, reload) are synthesised in designer/sfx.js

The user asked for Gyropup to talk like Moxie, with word bubbles and STT, and for cool sound effects. Adapting our own MIT Moxie code keeps one voice stack; the fallbacks keep the page working with nothing else running.

Targets are 20 g foam plates hinged at the bottom on posts; a hit knocks them over and the turret moves on (5 targets, pass = 60 % knocked down and 40 % of darts on target). Darts have a heavier rubber tip (0.55 g of 1 g) so they fly nose-first. The laser is off at the start and while slewing between targets, comes on when the turret is within 0.08 rad of its aim, and stays on for 0.35 s before the first shot. A 0.6 s pause follows each knock-down. The simulator's laser has AUTO / ON / OFF modes

The user said the darts looked cheesy (they tumbled and nothing reacted to a hit) and asked for the laser to come on only when getting ready to fire. A laser before every burst is also a sensible visible warning for kids.

designs/default.json runs from one DeWalt FLEXVOLT in 60 V mode (15S, 45–60 V) in a FLEXVOLT 60 V dock adapter. The parts that could not take 60 V are swapped: two Flipsky 75100 VESCs (14–84 V) drive the BotWheels instead of moteus (54 V max), a Mean Well DDR-30L-12 (18–75 V) makes the 12 V rail instead of the Pololu D36V50F12 (50 V max), and a Mean Well DDR-120C-24 (33.6–67.2 V) feeds the J501. A full pack sits at the RS-03's 60 V limit, so a regen clamp (brake chopper and 2 Ω resistor, ASSUMED part) trips at 58 V and the robot treats 58 V as full. The two-pack 40 V build is kept as designs/flexvolt-2x20v.json

The user has 60 V packs and asked for this build in the simulator. One pack is simpler to swap and dock than two in series. The designer now checks the clamp against every motor's limit; the build passes 9/9 scenarios at 17.0 kg (17.5 kg with the CAD part masses).

Where a fact was missing, a reasonable assumption is used and labelled as assumed: FLEXVOLT pack size about 157 × 92 × 110 mm; a WeAct STM32G474 balance controller; a pair of 20 W 12–30 V LED pods as headlights; turret pan/tilt torques (5 and 3 N·m) and a 40/60 pan/tilt mass split; darts of 1 g at 24 m/s with drag only

The user asked to make things up where needed and keep going until the design is realistic, fires darts in simulation and has forward lights.

Two FLEXVOLT packs in 20 V mode (5S3P each) wired in series give a 30–40 V bus. A Pololu D42V55F24 (24–60 V in) feeds the J501. Off-the-shelf adapters provide the latch and contacts; 10 AWG leads, a 60–80 A time-delay fuse at the battery for burst current, the adapters' blade fuses as backup. The robot enforces a 3.0 V/cell cut-off per pack and a current limit, because tool packs leave that to the tool

The user has 60 V FLEXVOLT packs and likes tool batteries (easy to swap and charge). Of the three ways to use them, the series pair is the only one with no problems: it passes all 7 simulated scenarios at 17.6 kg, with 153 min estimated runtime, 146 rpm leg speed and 8 km/h top speed. One pack in 20 V mode is too slow (73 rpm legs, 4 km/h); 60 V mode exceeds the moteus drivers' 54 V and the lidar converter's 50 V, and leaves no regeneration margin on the RS-03s.

The foam-dart turret (automated reload and firing) is developed in a separate lane. Gyropup defines and simulates a head payload interface (bolt pattern, power, data, mass and centre-of-mass height, recoil impulse) so any turret that meets it can mount

The turret is a large project with plenty of open-source work to build on; keeping an interface between the two lets both move independently. The user expects about 24 shots per magazine (drum or belt) and a trip to get a new magazine, so the dock doubles as the reload station.

A second design, about 5.8 kg, on the user's Waveshare DDSM115 hub motors (12–24 V, RS485) with Steadywin GIM6010-8 legs, a Raspberry Pi 5 and a Makita 18 V pack, for tuning the balance controller cheaply before the full robot

The DDSM115 is too weak and too low-voltage for the 15 kg robot (it catches only a 32° lean and exceeds its 24 V limit on 8S LiFePO4), but suits a small, safe-to-crash dev robot (55° catch, 42° slope in the model).

The VLP-16 hangs under the body between the legs; the legs and wheels block part of its view and that is accepted. The head is free for the turret

The user's call. Under the waist it is protected in falls and lowers the centre of mass; the depth cameras cover what the legs hide.

Gyropup drives outdoors and carries a foam-dart turret on its head for Tritium's neighborhood robot battles with kids. Kid safety is a hard requirement in hardware and software (dart energy limits, aiming limits, wireless e-stop, speed limits, pinch points)

The user's goal for the robot. Turret recoil and head mass, terrain and weather now drive the design and the simulation scenarios.

A low dock with spring contacts and a fiducial marker. Gyropup reverses over it using the rear depth camera and lidar, then crouches so a contact plate under the battery tray lands on the contacts; its weight holds them closed. The dock holds the charger (29.2 V for 8S LiFePO4); the robot enables its charge input only after it detects the dock

Legs make docking easy: lowering the body is precise and needs no connector alignment. Unattended charging is why LiFePO4 is preferred.

An 8S LiFePO4 pack (25.6 V nominal, 29.2 V full, about 20 V at cut-off) runs the J501, motors and drivers directly. It must deliver 60 A or more continuous; the pack (ready-made or custom high-rate cells with a 60–150 A BMS) is being sourced

LiFePO4 is safer to charge, which matters most because Gyropup will dock and charge itself unattended (D18). It is a preference, not a requirement ("If LiFeP04 doesn't work that's OK too"). 8S keeps the J501 inside 12–36 V; 10S would exceed 36 V. The cost is weight, about twice the mass per watt-hour of the Li-ion pack. The only LFP pack in the catalog so far (Bioenno, 10 A continuous) is a placeholder that fails the current check.

Waveshare 0.71 in DualEye LCD (two 160×160 round IPS screens, GC9D01, SPI, 11 g, $15.99) on the front of the head, driven by a small separate controller; no mouth

Expressive eyes carry most of a friendly robot's character; a separate controller keeps the character layer off the balance loop.

VLP-16 directly on top of the body (no mast), centred over the wheel axle, on rubber isolators, inside a printed crown that takes the hit in a fall. The mount stays a design variable (top, mast, under) for the loop to test

Moving 0.83 kg changes the centre of mass by at most about 16 mm, so view and falls decide. Under the body, the wheels and legs hide roughly a third of its view; on top it is the first part to hit the ground, so it needs a guard.

Gyropup lives in a private repository; everything generic (robot export, the project site, FEA glue, optimiser staging) goes into the public cad-agent, which stays modular for any CAD project

The user wants to build on a public toolkit while keeping this design closed, maybe opening it later.

Gyropup (gyro, because it balances itself; pup, because it is a small, friendly robot). It is its own robot, not a duck; Open Duck Mini was only a reference for charm and buildability

wobbit.com was taken. The user picked Gyropup from names with unregistered .com domains and kept it after finding a furry who uses the name.

MuJoCo for evaluation with a per-design LQR ported from Upkie; mjlab (from mjlab_upkie) for RL, Isaac Lab later; gmsh + scikit-fem FEA, CalculiX to verify; Optuna 5 with staged fidelity

The toolchain survey found these are the best maintained, permissively licensed options; no framework ships a wheel-legged scenario suite, so we write that part.

RobStride RS-03 leg actuators; ODrive BotWheel wheels with moteus n1 drivers

RS-03 is $225 for 60 N·m peak and 20 N·m rated with STEP files and torque–speed data; BotWheel has a fine encoder for slow balancing and is in stock in the US. The loop may change both.

Hip and knee motors stacked on the body with a four-bar link to the knee; legs move only fore and aft in v1

Diablo is built this way (DDT's sim model); motors on the body keep legs light; Diablo, Upkie and Ascento all balance without hip abduction.

The balance microcontroller drives the motors on its own CAN bus. The J501's built-in CAN-FD carries setpoints and state between the Jetson and the microcontroller, so no USB sits in the control path; the second J501 port is free for logging. The e-stop, a wireless relay and the controller's enable switch drive a contactor that cuts motor power while the computer stays up

Seeed's schematic shows two isolated CAN-FD ports on J25; the user wants the microcontroller to be the main CAN for the motors.

WREKD 7S2P Molicel P50B pack (29.4 V full, 252 Wh, 1.05 kg) runs the J501 directly; a 12 V regulator feeds the VLP-16

Keeps the J501 within 12–36 V without a converter; Diablo's motors are in the same class. The trade-off is that 48 V-rated leg motors run at about half speed.

Assumed for now: a WeAct STM32G474 core board (3× FDCAN, 170 MHz with FPU) with two CAN transceivers runs the 1 kHz loop and owns the motor CAN bus. Swap in one of the user's own boards when named

Needs a steady 1 kHz loop independent of Linux, CAN for the motors plus CAN to the J501 (2× CAN, or 1× CAN-FD shared), a UART for the VN-100, and an FPU. Waiting for the user's list of boards.

VectorNav VN-100 in its housing, mounted rigidly on the body near the centre of mass, read over serial at up to 800 Hz

A calibrated, temperature-compensated IMU the user has; the lidar mast flexes and carries a spinning rotor, so the IMU stays off it.

VLP-16 on top with a clear horizon; RealSense D435 facing forward and another facing back in a cradle that also fits the D415

GMSL is overkill for a robot this size, and the ZED X has no driver for the J501's MAX96724 GMSL board.

Jetson AGX Orin 32GB on the Seeed J501 in our printed enclosure, powered straight from the pack (12–36 V input)

It is what we have; a 7S pack stays under 36 V so no converter is needed.

Every structural part starts as PLA on a Bambu P1S; a material ladder (PLA → PETG → CF-filled) and SendCutSend metal or CF tube are upgrades with the same interfaces

Low cost and fast iteration. PLA softens at 54 °C, so parts touching motors start in PETG and the checks flag heat.

Design → CAD and robot model → simulation → loads and stress → manufacturing and cost → iterate without a human, with any components and any specs

The user's stated goal. Prior art has the pieces but nobody has published the whole loop, and stress checking is the missing step.

Two legs (hip + knee) ending in direct-drive wheels; it balances on the wheels and crouches, stands and hops on the legs

What the user asked for; the reference robots (Diablo, Ascento, Upkie) show the layout works.