free-humanoid-wheeled

Architecture

Free Humanoid Wheeled — an open hardware, open firmware, open data wheeled-base humanoid reference design. Sibling morphology to the bipedal Free Humanoid Platform. Designed for the shoal dock-B service mission. Shielded by the Free Humanoid Corpus.

Contents

A ~70 kg, paved-environment humanoid with a wheeled differential-drive base and the same upper body as the bipedal Free Humanoid Platform. Expressed as a single OpenLoco UDD descriptor that compiles to URDF, MJCF, STL meshes, BOM, and assembly through the existing OpenLoco toolchain. Every non-trivial design choice cites a specific entry in the Free Humanoid Corpus by id.

Phase 1 status (2026-05-07): descriptor v0.1.0 vendor-pinned to commercial-off-the-shelf components (T-Motor AK60-6 + mjbots Moteus n1 drive; Igus DryLin SHTS-15 tower lift; Livox MID-360 lidar; u-blox ZED-F9P RTK GNSS; VectorNav VN-100 IMU production / Adafruit BNO055 budget; Quectel RM502Q-AE 5G; NVIDIA Jetson Orin Nano Super + Raspberry Pi 4 Simplex compute; Bioenno 24V LiFePO4 battery). Wheeled-base BOM is ~$9–10.5k for the wheeled-base subsystem alone (excludes the inherited upper body). Phase 2 work begins with chassis fabrication + drive-base prototype + dry-side bench test. Dock-B mock cartridge-swap end-to-end Phase 3, open-water shoal dock-B municipal-utility pilot Phase 4.

This document is the canonical full-system spec for the wheeled morphology. It cross-references — and does not duplicate — the bipedal sibling’s ARCHITECTURE.md for everything inherited (upper body, sensing, compute, safety, manipulation, learning policy). Decisions that diverge from the bipedal sibling are argued in full here.


0. Design principles

Six principles, an extension of the bipedal sibling’s six (see Free Humanoid Platform ARCHITECTURE §0).

  1. Open everything. Hardware, firmware, descriptors, BOM, control software, simulation models, training data, documentation. CERN-OHL-S 2.0 / Apache 2.0 / CC-BY-SA 4.0 / CC0 1.0 by artifact class.

  2. Modular at the interface, not at the implementation. Inherited from the bipedal sibling. The morphology-swap is itself the worked example: the base subsystem is modular at the descriptor level — replacing the bipedal legs block with the wheeled base block produces this morphology, with the upper body unchanged.

  3. Mission-specific, not general-purpose. This platform is shaped to the shoal dock-B service mission. The bipedal sibling is shaped to the larger open question of “what is an open humanoid for human environments?” — that is the wider scope, this is the sharper one. A wheeled humanoid with no mission is a worse design than a bipedal humanoid with no mission, because the wheeled morphology gives up stair-traversal for a payload and runtime budget that only pays off when the mission demands it.

  4. shoal-integration-aware. The platform’s protocols (cartridge handshake, waste-hopper interface, dock biology-reservoir status) are designed against the shoal dock-B spec (shoal ARCHITECTURE §2.2, §3, §4). The integration is explicit — not hidden behind an abstract “tasks” API — because the integration is the platform’s reason to exist at v0.1.

  5. Fail-safe to inert. Inherited verbatim. Every subsystem has a defined fail-safe state. On the wheeled base specifically, the fail-safe is “drop to a stable wheel-locked stance, latch brakes, kill drive rails.” Slope-tilt and edge-proximity invariants are added to the bipedal sibling’s safety supervisor invariant set.

  6. Prior-art-shielded by deep corpus chain. Inherited. Every load-bearing decision in this document is shielded back to a corpus entry. The wheeled-base subsystem is shielded by the chain dlr-justin (2009) → pr2 (2010) → willow-pr1 (2008) → pepper-softbank (2014) → toyota-hsr (2012) → reachy-2-pollen-2023 (2023) → ascento (2019) → nao (2006), plus the fictional anchors forbidden-planet-robby (1956), magnus-robot-fighter (1963), b-9-lost-in-space (1965), silent-running-drones (1972).


1. Fork map: what we take from where

1.1 From Free Humanoid Platform (the bipedal sibling)

Everything above the wheeled base. Specifically:

Subsystem Inherited section in bipedal ARCHITECTURE What we take
Torso, arms, hands, head §2.2 Kinematic topology, §2.4 Actuators 21-DoF upper body topology: 1-DoF waist + 2-DoF neck + 2× 7-DoF arms + 2× 1-DoF underactuated synergy hands. Inherited verbatim.
Sensing §2.5 Sensing IMU at every major link, force-torque at each wrist, GelSight-style tactile at fingertips, head stereo camera. Plus wheeled-specific additions in §2.4 below.
Compute §2.6 Compute Joule SOM as recommended-default, with abstract HAL boundary. Verbatim.
Safety supervisor §2.7 Safety supervisor MathGround Simplex pattern. Verbatim, plus wheeled-specific invariants in §2.6 below.
Manipulation §2.9 Manipulation Underactuated 5-finger hand. Verbatim.
Learning policy §2.10 Learning policy Whole-body MPC + RL locomotion (now wheeled-locomotion-only) + IL manipulation. Verbatim above the locomotion layer.
Comms §2.11 Comms ROS 2, CAN-FD, gRPC, WebRTC. Verbatim, plus 5G + JANUS-receiver additions in §2.11 below.
End-of-life §2.13 End-of-life Same posture.
Governance README §Governance, CONTRIBUTING.md Verbatim.

The descriptor in descriptor/free-humanoid-wheeled.udd.json inherits the bipedal links/joints/actuator_slots for the upper body and replaces the leg blocks with the wheeled-base blocks.

1.2 From OpenLoco

The same UDD compiler. The morphology is humanoid_wheeled. OpenLoco already compiles humanoid_wheeled morphologies (Pepper-class wheeled-base mobile manipulators were among the original sixteen morphologies). The schema extensions added by the bipedal sibling for cycloidal / harmonic-drive / tendon-driven actuator types apply here unchanged.

1.3 From the Free Humanoid Corpus

The corpus is morphology-agnostic; this morphology cites a different overlapping subset than the bipedal sibling does. The new corpus chain leans heavily on the wheeled-base academic anchors:

Corpus id Year What it gives us
dlr-justin 2009 DLR Rollin’ Justin: the canonical academic disclosure of wheeled humanoid mobile manipulation with full impedance control. 17-year prior art on every aspect of this morphology family.
pr2 2010 Willow Garage PR2: omnidirectional wheeled mobile manipulation; the platform around which ROS was built. 16-year prior art.
willow-pr1 2008 Willow Garage PR1: cable-driven intrinsically-safe wheeled humanoid; 18-year prior art.
pepper-softbank 2014 SoftBank Pepper: wheeled-base humanoid social robot with omnidirectional drive.
nao 2006 SoftBank/Aldebaran NAO: not wheeled, but the SoftBank-Aldebaran lineage anchor that culminates in Pepper.
toyota-hsr 2012 Toyota Human Support Robot: telescoping-torso wheeled humanoid for domestic service. The closest existing prior art to this morphology’s design domain.
reachy-2-pollen-2023 2023 Pollen Robotics Reachy-2: open-source mobile humanoid with VR teleop. Anchors the open-source-wheeled-humanoid posture.
ascento 2019 EPFL Ascento: wheeled-balancing biped — anchor for any future wheeled-balancing variant of this morphology.
mit-cheetah-2 2014 MIT Cheetah 2: QDD actuator architecture; anchors the drive-motor controller choice.
mjbots-moteus 2019 mjbots Moteus: open BLDC controller; the recommended drive-motor controller.
forbidden-planet-robby 1956 Robby the Robot: 70-year fictional anchor for wheeled service humanoid with multi-language interface. The single most pointed fictional precedent for this morphology.
magnus-robot-fighter 1963 Magnus, Robot Fighter: 63-year fictional anchor for mass-produced humanoid civic deployment.
b-9-lost-in-space 1965 B-9 (Lost in Space): 61-year fictional anchor for civil-defense / service humanoid.
silent-running-drones 1972 Huey, Dewey, Louie: 54-year fictional anchor for compact service drones; the on-set drones were physically functional bipedal humanoid platforms.

The fictional chain is not decorative — forbidden-planet-robby in particular establishes a wheeled humanoid service robot with a multi-language interface in 1956. Any post-2010 patent claim that combines those three elements faces a 70-year-deep fictional anchor in addition to the academic chain.

1.4 From prior shoal architecture

The mission-integration spec (cartridge protocol, waste-hopper interface, biology-reservoir handshake) is taken directly from shoal ARCHITECTURE, specifically:

shoal section What the wheeled morphology inherits
§2.2 The gut cartridge Cartridge mechanical / fluid / electrical interface spec, used as the manipulation target for cartridge-rotation tasks
§3 The dock The dock-B (shore-tethered) topology; the platform is designed to dock-side at this configuration
§4 The protocols The dock-burst optical protocol; the platform receives status telemetry over this channel as a node on the dock LAN
§5 The cognitive layer shoal-fleet’s cartridge-rotation scheduling; the platform takes work orders from this layer
§9 Safety and ethics Indigenous and community consent for outdoor / public-space deployment; same posture

2. Platform spec: subsystem-by-subsystem

2.1 Form factor

Quantity Value Notes / shielding
Total system mass ~70 kg 30 kg base + 40 kg upper body. Lower than the typical wheeled-humanoid range (dlr-justin ~200 kg, pr2 ~225 kg, toyota-hsr ~37 kg) — Justin and PR2 are heavy because they’re research-grade with redundant compute and full omnidirectional drive; HSR is light because it has a single arm. The 70 kg target is shielded by the corpus median for wheeled service humanoids and is the mass at which a 30 kg payload at extended reach is statically stable on a 600×800 mm base.
Upper body height (standing) ~1.40 m Inherited from the bipedal sibling minus the leg length (the base + tower stand sits at 0.55 m above ground; the bipedal pelvis-to-head height is ~0.85 m).
Total height (standing) ~1.55 m Comparable to unitree-g1 (1.32 m), pal-talos (1.75 m), apptronik-apollo (1.73 m). At the small-shoulder-side of the human-scale band, intentional for indoor doorway clearance and for not visually overwhelming municipal-utility staff who share workspace with it.
Base footprint 600 mm wide × 800 mm long × 350 mm tall Wide stance for stability under upper-body manipulation loads. Clears standard 800 mm interior doorways. Comparable to pr2 (~668 mm), toyota-hsr (~430 mm — much smaller because single-arm), dlr-justin (~700 mm omnidirectional).
Sustained payload at extended reach 30 kg Possible because (a) wide base gives a large support polygon, (b) low-mounted battery lowers the COM, (c) no leg dynamics to budget. Shielded by dlr-justin (which routinely demonstrated ~10 kg per arm at extended reach with a much heavier base); 30 kg is the design-limit for a single shoal cartridge in transport plus a hopper-load of collected sediment.
Peak ground speed ~1.5 m/s Half-again the bipedal sibling’s 1.0 m/s. Below the 2 m/s threshold widely used for shared-space mobile-robot operation in service environments. Shielded by pr2 (~1.0 m/s), toyota-hsr (~0.8 m/s), pepper-softbank (~0.9 m/s).

2.2 Wheeled base

The morphology-distinguishing subsystem.

Drive geometry: differential-drive with two passive casters. Two large drive wheels at the rear (or center, see TBD below); two passive swivel casters at the front. This is the simplest wheeled topology that gives full planar mobility (excluding sideways translation) and is shielded by dlr-justin, pr2, pepper-softbank, nao (Nao’s later mobile-base derivatives), and the standard ROS-era mobile-manipulation chain.

TBD (architectural call): differential-drive vs. omnidirectional. Differential-drive (two drive wheels + casters, Pepper / Justin posture) is simpler, lower-cost, more robust on rough urban paving. Omnidirectional (Mecanum wheels or three-wheel Killough, PR2 posture) gives sideways translation, which is useful for dock-side cartridge swaps where the robot is parallel-parked next to a dock. Recommendation: differential-drive for v0; omnidirectional descriptor variant for environments where dock-side maneuvering volume is constrained.

TBD (architectural call): drive-wheel placement. Centered (two-wheel-balancing posture, requires active balance control — Ascento posture) gives a smaller turning radius. Rear-mounted with front casters (Justin posture) gives passive stability. Recommendation: rear-mounted with front casters for v0; the balancing variant is held as a future track shielded by ascento.

Wheel choice: 15” diameter pneumatic. Pneumatic for outdoor curb tolerance (a 15” pneumatic wheel will roll over a 50 mm curb edge without latching the safety supervisor). 15” is the smallest wheel size that gives that tolerance with a comfortable margin. Shielded by dlr-justin (which uses 200 mm hard-rubber wheels but on indoor surfaces only) and pr2 (which uses 800 mm-class drive wheels for the same curb-tolerance reason in outdoor variants). Solid-foam-filled tires are an alternative that trades curb tolerance for puncture-immunity.

Caster geometry: 8” diameter swivel casters. Mounted ~700 mm forward of the drive-wheel axle. Swivel free in azimuth, fixed pitch. Shielded by every wheeled-humanoid corpus entry — caster-front + drive-rear is the universal differential-drive pattern.

Battery placement: low-mounted, between the drive wheels. Two ~1 kWh modules, each ~20 kg. Mounting between the drive wheels lowers the COM to ~150 mm above the ground, which gives the static-stability budget for the 30 kg sustained payload at extended reach.

Drive motors and controllers: QDD direct-drive into 15” wheel hub, with planetary reduction. Each drive wheel runs a single BLDC motor with planetary reduction (target ratio ~10:1). Controller: mjbots Moteus n1 (one per wheel). Shielded by mit-cheetah-2, mini-cheetah, mjbots-moteus. The drive-motor torque target is ~80 Nm continuous at the wheel, ~250 Nm peak — sufficient to climb a 10° ramp at 1 m/s with full payload.

Encoder placement: at motor shaft. The reduction is between motor and wheel, so motor-shaft encoder gives the cleanest position feedback. Wheel-axle encoder optional for slip detection on outdoor surfaces.

Tower stand: rigid aluminum tower mounting the upper body’s pelvis at 0.55 m above ground. Replaces the bipedal sibling’s leg stack. The tower includes a single 1-DoF prismatic “torso lift” actuator giving ±150 mm of vertical reach (compressed to extended), shielded by toyota-hsr’s telescoping-torso disclosure. With the torso lift extended, the head sits at ~1.70 m; with it retracted, ~1.40 m.

Drive-wheel suspension: TBD. Hard-mount (Justin posture, no suspension) is mechanically simplest but transmits curb shocks to the upper body. Independent-arm + spring (PR2 posture) absorbs shocks at the cost of complexity. Recommendation: hard-mount for v0; spring-arm variant in a future descriptor variant.

2.3 Upper body

Inherited from Free Humanoid Platform ARCHITECTURE §2.2 (kinematic topology), §2.4 (actuators), §2.5 (sensing), §2.9 (manipulation), §2.10 (learning policy). 21 DoF, hybrid actuator distribution (harmonic-drive at shoulder/elbow, tendon at wrist/hand, QDD at neck/waist), underactuated 5-finger hands, GelSight tactile fingertips, head stereo camera, IMU at every major link.

The corpus citations for the upper body are unchanged from the bipedal sibling and are recapitulated in prior-art/INDEX.md for indexing convenience — but the canonical statement of those citations is the bipedal sibling’s prior-art/INDEX.md.

2.4 Sensing additions over the bipedal sibling

In addition to the inherited bipedal sensing stack, the wheeled morphology adds:

2.5 Compute

Same Joule SOM commitment as the bipedal sibling (see Free Humanoid Platform ARCHITECTURE §2.6). The compute distribution is unchanged: safety supervisor on a separate compute domain from the high-performance controller (Simplex pattern); perception runtime on its own compute domain.

2.6 Safety supervisor

Inherits the bipedal sibling’s MathGround pattern verbatim, plus three wheeled-base-specific invariants:

  1. Slope tilt limit. Base pitch and roll are continuously monitored. If pitch exceeds 12° or roll exceeds 8° (inherited from dlr-justin’s reported tilt thresholds), the supervisor latches the brakes, drives the upper body to a low-COM safe pose (arms tucked, torso lift retracted), and reports a slope violation upstream. The thresholds are static-stability limits with the 30 kg payload at extended reach; reducing payload extends them.

  2. Edge-proximity hard-stop. The lidar is processed continuously for ground-plane discontinuities (cliff or edge ahead of the drive wheels). If a discontinuity > 80 mm vertical drop is detected within the projected stopping distance, the supervisor latches the brakes immediately. This is the “do not drive off the dock edge” invariant. Shielded by dlr-justin, pr2, and the universal mobile-robot cliff-sensor pattern.

  3. Low-battery-on-shore-power latch. When the platform is on shore power (umbilical to the dock-B charging station) and battery falls below 20%, the supervisor refuses any motion command that would un-tether the platform. This prevents the platform from “leaving the charger before charged enough to return” — a failure mode reported in pr2 and toyota-hsr operational logs.

The shielding chain for the safety supervisor as a whole is unchanged from the bipedal sibling: asimov-positronic-robots (1940) → williamson-folded-hands (1947) → forbidden-planet-robby (1956) → magnus-robot-fighter (1963) → b-9-lost-in-space (1965) → hal-9000 (1968) → silent-running-drones (1972) → asimovs-zeroth-law (1985) → data-tng (1987) → robocop-1987 (1987) → sherman-simplex-architecture (1995) → reachability-analysis-safe-control (2005) → iso-10218-collaborative-robots (2006) → control-barrier-functions (2007) → runtime-assurance-rta (2010) → shielding-rl (2018). The fictional chain is denser at the front in this morphology — forbidden-planet-robby, magnus-robot-fighter, b-9-lost-in-space, silent-running-drones all directly anticipate civic-service / civil-defense / hazard-warning autonomous behavior, which is exactly the deployment context for a dock-B service humanoid.

2.7 Manipulation

Inherited from Free Humanoid Platform ARCHITECTURE §2.9. 5-finger underactuated synergy-driven hand, GelSight tactile fingertips, force-torque at the wrist. Corpus: pisa-iit-softhand, shadow-dexterous-hand, dlr-hand-ii.

The shoal-mission-specific manipulation targets — cartridge handle, hopper handle, dock-side maintenance fixtures — are designed to be compatible with the underactuated 5-finger hand without specialized end-effector swaps. This is a constraint on shoal’s dock-B mechanical design, contributed back upstream into shoal ARCHITECTURE §2.2.

2.8 Locomotion control

The locomotion-control stack is the largest divergence from the bipedal sibling.

Low level: differential-drive kinematics + PID per wheel. Standard textbook differential-drive (state: x, y, heading, linear and angular velocity; control: left and right wheel torque). Shielded by pr2, dlr-justin, the entire ROS mobile-base lineage, and the textbook differential-drive treatments going back to the 1980s.

Mid level: differential-drive MPC. Model-predictive control with horizon ~1 s, cycle ~50 Hz. Constraints: linear velocity ≤ 1.5 m/s, angular velocity ≤ 1.0 rad/s, slope-tilt and edge-proximity safety-supervisor handoffs. Cost: tracking-error + smoothness + obstacle-avoidance penalty. Shielded by the broad MPC-on-mobile-robots literature.

High level: behavior tree (via OpenLoco’s SkillGraphRuntime), with three primary modes:

  1. Free-roam: outdoor obstacle-avoidance navigation, lidar-based, GPS-aided. Used for transit between dock-B sites within a fenced facility.
  2. Dock approach: fine-positioning relative to the dock’s optical fiducial, using head stereo camera + dock-burst optical handshake (shoal ARCHITECTURE §4). Targets ±20 mm positional accuracy at the cartridge port.
  3. Manipulation static: wheels locked, brakes latched, base treated as immobile foundation for upper-body MPC. The bipedal sibling’s §2.10 whole-body MPC reduces to upper-body-only MPC in this mode.

Curbedge detection. A dedicated subroutine in the perception domain processes the lidar’s ground plane to flag curb edges, gutter drops, and dock edges. Output is a binary “edge ahead” signal subscribed by the safety supervisor for the edge-proximity hard-stop invariant (§2.6 #2). Shielded by dlr-justin and the universal mobile-robot cliff-sensor pattern.

2.9 Mission integration with shoal

This is the load-bearing original engineering for the wheeled morphology and is enumerated in §5 below.

The integration surface, shielded only weakly by the corpus (because the cross-product of wheeled-humanoid + microbial-water-remediation-fleet is novel):

2.10 Power

Component Spec Notes
Primary battery 2 kWh total, 48 V nominal Two ~1 kWh modules, mounted between the drive wheels (low COM). Doubled from the bipedal sibling’s 1 kWh; possible because no leg cavity constraint. Estimated runtime: 4–8 hours of dock-B service depending on duty cycle.
UPS 0.2 kWh, 24 V nominal Mounted high in the tower stand. Dedicated to compute, sensors, and the safety supervisor; isolated from the drive bus. Holds the safety-supervisor invariant hot for ≥ 20 minutes after a primary-battery fault, sufficient to drive the platform to a stable wheel-locked stance and report.
Hot-swap interface Inherited from the bipedal sibling’s commitment #6 Same CC0 commons spec contributed to OpenLoco UDD. The dual-module geometry of the wheeled base means hot-swap is single-module-at-a-time — one module remains live during swap.
Shore-power umbilical 48 V DC over locking connector at dock-B The dock-B “hardwired power” channel from shoal ARCHITECTURE §3. Connector geometry TBD; preferred is the same connector as the bipedal sibling’s hot-swap interface.

Shielded by hyundai-boston-dynamics-spot, unitree-h1, apptronik-apollo, tesla-optimus. The doubled capacity and the shore-power umbilical are the only divergences from the bipedal sibling.

2.11 Comms

Inherited from the bipedal sibling’s §2.11. On-robot CAN-FD; off-robot ROS 2, gRPC, WebRTC. Plus three additions for the wheeled morphology:

2.12 End-of-life

Same posture as the bipedal sibling. See Free Humanoid Platform ARCHITECTURE §2.13.

The wheeled-base-specific end-of-life concerns:


3. The descriptor

The single source of truth is descriptor/free-humanoid-wheeled.udd.json. It is OpenLoco UDD-compliant. The OpenLoco compiler produces:

To compile:

cargo run -- validate /path/to/free-humanoid-wheeled.udd.json
cargo run -- generate /path/to/free-humanoid-wheeled.udd.json --all --output-dir out/
cargo run -- bake /path/to/free-humanoid-wheeled.udd.json --output-dir out/meshes/

The descriptor structure:

The descriptor currently uses placeholder mass / inertia values. Iterating these to physical-build accuracy is a load-bearing follow-up task, gated on the wheeled-base architectural-call resolution (§2.2 TBDs).


4. Repository structure

free-humanoid-wheeled/
  README.md
  ARCHITECTURE.md          (this document)
  CONTRIBUTING.md
  LICENSE-HARDWARE         CERN-OHL-S 2.0
  LICENSE-SOFTWARE         Apache 2.0
  LICENSE-DOCS             CC-BY-SA 4.0
  LICENSE-DATA             CC0 1.0

  descriptor/
    free-humanoid-wheeled.udd.json   the source of truth
    README.md

  base/                    wheeled-base subsystem
  upper-body/              inherited from free-humanoid-platform
  electronics/             PCB pointers, power tree, Joule SOM
  firmware/                HAL, behaviors, MathGround, comms
  control/                 differential-drive MPC, manipulation policy
  sim/                     MuJoCo / Gazebo / Drake model pointers
  docs/                    additional design notes

  prior-art/
    INDEX.md               corpus entries, by subsystem

5. What needs original engineering

Not covered by existing prior art and requiring genuine new disclosure:

  1. The dock-B service-mission protocol stack. The cross-product of wheeled humanoid + microbial water-remediation cartridge fleet + municipal water-utility EAM integration is novel. Specifically: (a) the cartridge-rotation handoff between platform and dock; (b) the waste-hopper retrieval protocol; (c) the dock biology-reservoir status handshake; (d) the JANUS-receiver-as-coordination-hub pattern; (e) the EAM REST API. This is the work that is genuinely new and that the rest of this scaffold is designed to support.

  2. The wheeled-base UDD descriptor. OpenLoco already supports wheeled morphologies, but the humanoid_wheeled descriptor for this specific Justin-class differential-drive + tower-stand + bipedal-upper-body composition is a new artifact in the OpenLoco corpus.

  3. The shoal cartridge / hand mating geometry. The cartridge handle and hopper handle in shoal ARCHITECTURE §2.2 need to be jointly designed with the underactuated 5-finger hand from the bipedal sibling. This is mechanical-design work distributed between the two repos.

  4. The MathGround integration shim for the wheeled supervisor invariants. The slope-tilt, edge-proximity, and shore-power-latch invariants of §2.6 are wheeled-specific and need formal specification + verification in the MathGround framework. The bipedal sibling’s MathGround integration shim is the starting point.

Everything else is fork, port, or integration of existing open work, with corpus citations.


6. Roadmap

Phase 0 — months 0–3: scaffold and corpus citation

Phase 1 — months 3–9: simulation-buildable

Phase 2 — months 9–18: dock-side prototype

Phase 3 — months 18–30: first municipal pilot

Phase 4 — months 30+: ecosystem


7. Safety, ethics, and what we won’t do

Same posture as the bipedal sibling. See Free Humanoid Platform ARCHITECTURE §7.

Wheeled-morphology-specific:


8. Why open

Same three reasons as the bipedal sibling. See Free Humanoid Platform ARCHITECTURE §8. Trust, distribution, compounding.

The wheeled-morphology-specific argument: municipal water-utility deployments are exactly the domain where closed-source physical AI is least acceptable. A water utility cannot deploy a closed system to maintain shoal docks at public stormwater outfalls — the regulatory, community-consent, and public-records postures all push against it. An open platform is the only deployment vehicle that fits the regulatory and community context of municipal-utility deployment at scale.


9. Architectural commitments

The eight load-bearing decisions for this morphology, resolved. Each commitment is shielded by the cited prior-art chain.

# Decision Commitment Shielding chain (corpus ids)
1 Base topology Differential-drive with two large rear-mounted drive wheels and two passive front swivel casters. Omnidirectional and wheeled-balancing variants held as future descriptor variants. dlr-justin (2009, wheeled humanoid mobile manipulation), pr2 (2010, omnidirectional drive — held as future variant), pepper-softbank (2014, wheeled-base humanoid social robot), toyota-hsr (2012, wheeled domestic service humanoid), nao (2006, SoftBank-Aldebaran lineage anchor), ascento (2019, wheeled-balancing — held as future variant). Plus fictional anchors: forbidden-planet-robby (1956, wheeled service humanoid with multi-language interface — 70-year prior art), b-9-lost-in-space (1965, civil-defense / hazard-warning humanoid — 61-year prior art), magnus-robot-fighter (1963, mass-produced civic-deployment humanoid — 63-year prior art).
2 Mass budget ~70 kg total system. 30 kg base (battery + drive + casters + tower stand) + 40 kg upper body (the bipedal sibling’s 50 kg minus its leg mass). dlr-justin (~200 kg, research-grade), pr2 (~225 kg, research-grade), toyota-hsr (~37 kg, single-arm), pepper-softbank (~28 kg, single-arm). The 70 kg target is the corpus median for a dual-arm wheeled service humanoid and is the mass at which a 30 kg payload at extended reach is statically stable on a 600×800 mm base.
3 Drive-wheel choice 2× 15” diameter pneumatic drive wheels with QDD direct-drive into the wheel hub via ~10:1 planetary reduction. mjbots Moteus n1 controllers. mit-cheetah-2 (2014, QDD architecture), mini-cheetah (2019), mjbots-moteus (2019, open BLDC controller), dlr-justin (curb-tolerance precedent), pr2 (large drive-wheel precedent for outdoor variants). Pneumatic + 15” is the smallest configuration that gives 50 mm curb tolerance with comfortable margin.
4 Battery and runtime 2 kWh primary + 0.2 kWh UPS. Two 1 kWh modules between the drive wheels for low COM. Hot-swap is single-module-at-a-time. hyundai-boston-dynamics-spot (hot-swap legged platform), unitree-h1, apptronik-apollo, tesla-optimus. Doubled capacity over the bipedal sibling possible because no leg cavity constraint. The CC0 hot-swap interface spec from the bipedal sibling’s commitment #6 carries over.
5 Tower-stand torso lift 1-DoF prismatic torso lift, ±150 mm range. Mounted at the top of the rigid aluminum tower stand. Replaces the bipedal sibling’s leg stack. toyota-hsr (2012, telescoping-torso wheeled humanoid). HSR’s telescoping-torso pattern is the single most direct prior-art anchor for this commitment.
6 Upper-body inheritance Inherit the bipedal sibling’s upper body verbatim. 21-DoF (1-DoF waist + 2-DoF neck + 2× 7-DoF arms + 2× 1-DoF underactuated synergy hands). Hybrid actuator distribution: harmonic-drive at shoulder/elbow, tendon at wrist/hand, QDD at neck/waist. All of the bipedal sibling’s §9 commitments #1, #4, #8, #10. The inheritance is governance-load-bearing: a divergence in the upper body would fork the descriptor and the corpus citations, which is exactly what the OpenLoco substrate-and-tenant pattern is designed to avoid.
7 Outdoor sensing additions Add 16-channel rotating lidar (tower-top), GPS, 5G modem, JANUS acoustic-receiver, dedicated 2-DoF base tilt sensor. dlr-justin (lidar + tilt-supervised drive), pr2 (lidar additions post-2012), the broader mobile-manipulation lineage; JANUS-receiver is shielded by shoal ARCHITECTURE §4. Cellular and GPS modules are commodity and not patent-thicket-dense.
8 Safety supervisor wheeled invariants Add three invariants to the bipedal sibling’s MathGround Simplex pattern: slope tilt limit (12° pitch / 8° roll), edge-proximity hard-stop (80 mm vertical drop), low-battery-on-shore-power latch. The bipedal sibling’s full safety chain (asimov-positronic-robots 1940 → sherman-simplex-architecture 1995 → control-barrier-functions 2007 → runtime-assurance-rta 2010 → shielding-rl 2018) plus dlr-justin for the slope-tilt invariant; plus pr2 and toyota-hsr for the low-battery latch. The fictional layer is denser at the front (forbidden-planet-robby, magnus-robot-fighter, b-9-lost-in-space, silent-running-drones) for civic-deployed service humanoids.
9 Mission integration Designed against the shoal dock-B service mission as the v0.1 deployment target. Cartridge-swap, waste-hopper retrieval, biology-reservoir handshake, and water-utility EAM integration are first-class subsystems. shoal ARCHITECTURE §2.2 (cartridge), §3 (dock), §4 (protocols), §5 (cognitive layer). The cross-product (wheeled humanoid + microbial water-remediation fleet + municipal-utility EAM) is the platform’s load-bearing original engineering and is enumerated in §5 as such.
10 Reference design vs. specific build Reference design — the descriptor (free-humanoid-wheeled.udd.json, CC0) is the canonical artifact. Specific physical builds with chosen vendor parts live in tenant repos. Same posture as the bipedal sibling. The bipedal sibling’s §9 commitment #9. Matches OpenLoco’s substrate-vs-tenant pattern.

These commitments are this document’s defaults. Subsystem sections (§2) reflect them in detail. Future amendments require a corpus-citation update demonstrating that any new shielding chain is at least as deep as the chain it replaces.


Free Humanoid Wheeled — sibling morphology to the Free Humanoid Platform — designed for the shoal dock-B service mission — Phase 1 v0.1.0 — 2026-05-07.