Bodies should be cheap

One water-hydraulic source pressurises the whole machine. Hundreds of low-mass fibres pull on the joints, metered by valves that cost very little. That substrate ships as BoidKit; Limboid and Limboid Centaur are built on top of it.

Abstract

A motor, its reduction stage, bearings, sensing, drive electronics and a housing: that assembly repeats at every joint of a conventional machine, thirty-odd times over, and cost tracks the repetition almost linearly. Limboid amortises the expensive parts instead — one pressure source, one pack, one compute node, one cooling loop. What is left at an axis is a valve, two lines, a fibre bundle and a sensor. Development state is labelled per claim below; §07 lists the results that would falsify the architecture.

Get the build log Read the architecture

Shared pressure· Parallel muscle· Deterministic control Rev. 2026.08

§ 01

Reality is the bottleneck

Reasoning got cheap. Putting a hand somewhere reliably, repeatedly, and knowing afterwards that it arrived, did not.

What obstructs this is mostly mechanical and largely arithmetic. Replicate a motor, reduction stage, bearings, sensing, drive electronics and structure across thirty axes and both mass and cost follow the axis count; failure modes multiply rather faster than that. Bodies stay scarce for roughly the reason mainframes did.

Limboid moves the expense inward. Pressure generation, energy storage, compute and cooling are shared across the machine, which leaves at each axis a valve, two lines, a bundle of fibres, some geometry and a sensor — parts whose marginal cost is counted in units rather than hundreds.

Claim 1

Centralised fluid power, distributed lightweight actuation

A single water-hydraulic source pressurises the machine, manifolds distribute it, and bundles of low-mass fibres pull on each joint. Load divides across the bundle, so nothing out in the limb has to be individually large.
Claim 2

Cheap deterministic control; expensive intelligence only where it earns its place

Pressure, position and contact loops close on microcontrollers at 100–200 Hz. Estimation, whole-body control, perception and policy sit above that on Linux-class compute. Large models advise. Nothing with variable latency touches the plant.
Claim 3

Embodiment is an information problem

Calibration drift, pose uncertainty, replay and self-modelling belong inside the learning objective. Handing them to a human as setup work is how a machine ends up confidently wrong about where its own gripper is.
Claim 4

Manufacturing economics is a system constraint

Commodity motors and electronics; fluidic and structural parts off digital-fabrication equipment. Precision metal survives where physics leaves no alternative — shafts, wear surfaces, a handful of fittings.

§ 02

The stack

Four planes. Everything further down this page zooms into one row, or into an edge between two of them.

BoidKit

physical actuation substrate

Pump cassettes & accumulator
Pressure / return trunks
Zone manifolds
Progressive proportional valves
Parallel artificial-muscle bundles
Water / glycol working fluid

Body

morphology · sensing · compute

Humanoid & centaur topologies
Underactuated tendon hands
Cameras · IMU · encoders
Contact & pressure sensing
Dual CAN-FD trunks
Independent safety supervisor

Runtime

deterministic control

Zone joint loops · 100–200 Hz
Whole-body control · 20–100 Hz
State estimation
Explicit calibration state
Flow & pressure budgeting
Morphology graph

Cognition

planning · learning · memory

Persistent world & self model
Active information gathering
Skills & learned policies
Event-sourced causal memory
Verification before belief
Replay & audit

L0 to L1 is where the product boundary sits. Take the substrate and put your own body, runtime and policy above it, or take the finished machine.

§ 03

BoidKit

Actuation and body construction, sold as an architecture rather than a finished bill of materials — which is also how it reaches machines we would never have thought to build.

Table 1 — Component inventory and development state
Component	State	Function	Key parameters
Hydraulic artificial muscle	Active research	Linear contractile fluid actuator. Bundled in parallel, so each fibre carries a fraction of joint load and none of them individually has to be large.	4–10 fibres / side hundreds / body
Progressive proportional valve	Prototype	One 4/3 valve meters pressure and return across an antagonistic joint. Port geometry opens progressively: fine resolution off null, large area at full rotation.	pressure-balanced 4/3 low-torque rotary servo
Zone manifold	Prototype	Galleries, valve seats, sensor ports and mounting, collapsed into one digitally fabricated unibody and parameterised off the morphology description.	printed body reamed & lapped seats
Pressure core	Engineering target	Two cassettes merge through check valves into an accumulator. Behind them sits a guard carrying relief, sensing and a normally-open dump path that bounds stored energy.	40–60 bar 8–14 L/min return < 2 bar
Working fluid	Engineering target	Water, or water and glycol, in place of petroleum oil. Cleanup indoors gets easier and proof testing gets much safer; seal chemistry and corrosion get harder.	costs: corrosion, lubricity, microbial growth, cavitation
Zone controller	Engineering target	Owns encoder sampling, valve zero, calibration maps, setpoint interpolation, limiting, fault detection, and what happens when an upstream packet arrives late.	100–200 Hz dual CAN-FD
Underactuated hand	Engineering target	Four electric tendon drives: two opposing finger groups, thumb flexion, and thumb opposition carrying the transition between pinch and power grasp.	4 drives / hand coupled distal joints
Backscatter proprioception	Active research	The fluid network doubles as a sensing medium: a root transducer injects a pressure chirp, and every branch modulates its own reflection on the way back.	delay identifies branch

CoreBattery · pumps · accumulatorheavy, one of them

TrunkPressure & return lineslight, routed like vasculature

ZoneManifold + proportional valvescheap, many of them

AxisAntagonistic muscle bundlesalmost massless at the joint

Fig. 1Power leaves one heavy core and arrives at many near-massless endpoints. The separation is the reason for using fluid power; it is not evidence that hydraulics are cheap.

What it does not buy

Seals, leakage, pressure loss, contamination, thermal management, valve precision, pump efficiency, fatigue: none of these gets easier here, and several get worse once oil is replaced by water. Shared pressure buys a different allocation of cost and mass, which is a narrower claim than it sounds. Whether that allocation survives 10⁵–10⁷ cycles is Q1 and Q6 in §07, and the honest position today is that nobody here knows.

Full breakdown — BoidKit

§ 04

The machines

Two topologies over one substrate. Doors, counters, shelves, tools and handoffs fix the upper-body geometry; human anatomy does not.

Limboid Centaur

quadrupedal base · humanoid upper body
Engineering target

Legs: 4 × ~3 hydraulic axes
Torso / neck: 2 + 2 axes
Arms: 2 × ~7 hydraulic axes
Hands: 2 × 4 electric tendon drives
Total: ~30 hydraulic + 8 electric
Support polygon: 4 contacts, static

Four contacts hold a large support polygon while the arms work, which is why this is the platform going first — ahead of the much harder balance problem, and with payload and control burden both improving in the meantime. The word describes topology and nothing else. A tail, fur or hooves would each cost mass and buy nothing.

Limboid

bipedal humanoid
Long-term vision

Legs: 2 × ~6 hydraulic axes
Torso / neck: 2 + 2 axes
Arms: 2 × ~7 hydraulic axes
Hands: 2 × 4 electric tendon drives
Total: ~30 hydraulic + 8 electric
Support polygon: 2 contacts, dynamic

Narrowest footprint, widest set of human spaces, fewest rooms needing modification — and, by a wide margin, the harder control problem. It arrives through the centaur rather than instead of it. Substrate, runtime and hands are shared; what differs between them is a graph.

Fig. 2Limboid Centaur kinematic topology. Amber marks the fluid core and the hands.

Fig. 3Limboid bipedal topology. Substrate and hands are unchanged; the difference lives in the morphology graph.

Table 2 — Shared systems
System	State	Specification
Power	Engineering target	16s LiFePO₄ · 51.2 V nominal · 58.4 V full · 32 Ah ≈ 1.64 kWh Mains stays offboard. Isolated low-voltage DC enters the chassis through an interlocked inlet, then BMS, service disconnect, precharge and main contactor in that order.
Control rates	Engineering target	joint 100–200 Hz · whole-body 20–100 Hz · vision ~30 Hz · planning async Starting bands. Hydraulic bandwidth, valve geometry, sensor latency, flexible modes and contact regime will each move them. Remote reasoning never enters a deadline.
Safety	Engineering target	inhibit pump torque · open dump path · open main contactor A supervisor independent of main compute holds authority over the energetic subsystems, with relief valve, fuses, check valves, load-hold geometry and mechanical stops underneath it.
Qualification	Engineering target	fabricate → inspect → hydrostatic proof → decay → cycle → calibrate → safe-stop → release Every self-fabricated fluid-power part clears all eight gates first. Fabricating at home changes the equipment, not the pressure-vessel engineering.
Face	Long-term vision	display, not mechanism Gaze and expression become rendering problems, status becomes legible at a distance, and whether the machine is autonomous or driven stops being a guess.

Full breakdown — the machines

§ 05

Reality is not an API

Software agents receive a world already discretised into click, call, run. Expand “pick up the object” and what comes out is camera calibration, segmentation, pose uncertainty, two coordinate transforms, trajectory generation, backlash, contact detection, friction, grasp closure, load verification, and a recovery branch for slip. Constructing the measurement is most of the task.

perceive calibrate / model plan act verify remember improve

Calibration state is state

Intrinsics, extrinsics, tool-centre transforms, encoder zeros, valve deadband, pressure curves, contact bias, timing offsets — each carried explicitly, each with a covariance beside it. A planner can then ask which uncertainty is binding: the world's, or its own.

calibration:
  head_camera_extrinsic: { estimate, covariance, observed_at }
  left_gripper_tcp:      { estimate, covariance }
  valve_zero:            { zone_2_axis_4 }
  time_sync_error_ms:    …

Causal agency, not fluent narration

Alignment between the internal transition model and the world's is bought through intervention. Observation alone will not close it. Two trajectory-level quantities carry most of the signal.

IG (a) = H (Θ ∣ h_{t}) - E_{o^{'}} H (Θ ∣ h_{t}, a, o^{'})

(1)

ρ_{τ} = \frac{I (A_{t}; O_{t + τ} ∣ O_{\leq t})}{H (O_{t + τ} ∣ O_{\leq t})}

(2)

Information gain in (1); self-effect in (2) — the degree to which acting now determines what gets observed at t+τ. Neither is a finished objective. Both point the same way: the machine is rewarded for making reality more predictable by touching it.

Full breakdown — research threads

§ 06

Where we actually are

A dependency order. Dates are absent because they would be invented.

Phase 0

Characterise primitives

muscle force / contraction / fatigue rigs · valve deadband, leakage, flow curves · pump efficiency · material compatibility · proof and burst testing

Active

Phase 1

Single-axis closed loop

antagonistic bundle · progressive valve · encoder · 100–200 Hz local controller · disturbance and load-hold tests

Phase 2

Limb / multi-axis manifold

shared galleries · synchronised axes · flow starvation · thermal and leak testing

Queued

Phase 3

Quadruped base

static and quasi-static gait · load carriage · contact estimation · safe stop under line and pump faults

Queued

Phase 4

Torso & manipulation

torso and neck · one arm, then two · underactuated hands · visual and tactile grasping

Queued

Phase 5

Autonomous reality contact

persistent world and self model · calibration-aware perception · event-sourced memory · task autonomy with verification

Queued

Phase 6

Bipedal Limboid

the humanoid topology on a proven substrate, runtime and hand

Queued

Phase 7

Scale & manufacturing

design for assembly · cycle-life distributions · field repair · quantified BOM at volume · external developer interfaces

Queued

Table 3 — Programme lineage
Intensive R&D	early 2023 → mid 2024, ~18 months
Valve & muscle prototypes	100+
Major actuation redesigns	several, not one render
Also built	parameterised unibody manifold · custom driver PCBs · muscle/valve simulation
Original thesis	one shared prime mover, many low-cost fluidic actuators

The thesis has not moved since 2023. Scope has: an actuation kit became a machine, with manipulation, autonomy and fabrication attached to it along the way.

On numbers. Earlier public material from this programme carried per-DOF and whole-robot figures in the low hundreds of dollars. Those were exploratory, never measured, and they are gone from this site rather than caveated in a footnote.

A sub-$1,000 machine at manufacturing scale remains the economic target. Stated as a target every time it is stated at all.

What replaces them

$ / controlled axis
kg / controlled axis
joint torque, cont. & peak
actuator specific force
hydraulic efficiency vs load
pump efficiency map
valve leakage & deadband
energy per metre
energy per manipulation
payload / robot mass
grasp success under variation
mean repair time
parts count
fabricated vs purchased BOM
cycle-life distributions
BOM at 1 / 100 / 10k units

§ 07

What is blocking us

Unresolved, and published for that reason. Anything that gets answered leaves this list and turns into a table.

Which muscle geometry maximises lifetime-adjusted force per dollar, as against peak force on a fresh sample?

Do water-hydraulic seals and valves stay low-leakage over useful cycle counts with mostly polymer components?

At what pressure is total system mass minimised, once pump, line, valve, actuator and structural scaling are all counted together?

Can one progressive rotary valve deliver fine low-flow control and high peak flow from cheap servo torque?

Which fabrication and finishing process yields acceptable leakage and deadband repeatably, rather than on a good day?

Which axes can go passive or compliant before manipulation suffers for it?

Are four drives per hand enough for the everyday object distribution, and what specifically demands a fifth?

Can static joint load be held mechanically, so that standing still costs no hydraulic power?

What joint-loop bandwidth is reachable with compliant fibres and low-cost valves in the loop?

Q10

How much pressure sensing survives removal, given good position and contact data?

Q11

Can individual backscatter branch responses be separated while the pump is running?

Q12

Which physical skills survive fleet replay without catastrophic sim-to-real brittleness?

Appendices

Deeper

§ 08

The build log

Test plots, failures, CAD revisions, BOM changes, the occasional working joint. Sent when something measured exists, and otherwise not sent.

Bodies should be cheap.

Centralised fluid power, distributed lightweight actuation

Cheap deterministic control; expensive intelligence only where it earns its place

Embodiment is an information problem

Manufacturing economics is a system constraint

BoidKit

Body

Runtime

Cognition

What it does not buy

Limboid Centaur

Limboid

Calibration state is state

Causal agency, not fluent narration

Characterise primitives

Single-axis closed loop

Limb / multi-axis manifold

Quadruped base

Torso & manipulation

Autonomous reality contact

Bipedal Limboid

Scale & manufacturing

Bodies should be cheap