Low-cost embodied robotics
One shared water-hydraulic core drives hundreds of lightweight artificial muscles through cheap proportional valves. We ship that substrate as BoidKit for developers, and we build Limboid — the humanoid and centaur robots that run on it.
Machine intelligence is no longer the scarce input. Reliable physical agency is. A model that can reason but cannot establish trustworthy causal contact with the world is not yet doing useful physical work.
The thing standing in the way is mostly mechanical and mostly economic. Conventional robots replicate a motor, a reduction stage, bearings, sensing, drive electronics and a structural housing at every single joint. Cost, mass and failure modes scale with the number of degrees of freedom. So bodies stay rare, and everything downstream of a body stays rare with them.
Limboid attacks the marginal cost of a controlled degree of freedom. Expensive resources — pressure generation, energy storage, compute, cooling — are centralised and amortised across the whole machine. What remains at each axis is close to a valve, two lines, a bundle of fibres, some geometry and a sensor.
A shared water-hydraulic pressure source feeds many low-mass artificial-muscle bundles through compact valve manifolds. Load is spread across parallel fibres instead of one expensive actuator per joint.
Microcontrollers close pressure, position and contact loops. Linux-class compute handles estimation, whole-body control, perception and policy. Large models advise; they are never in the hard real-time path.
Calibration, uncertainty, replay and self-modelling are first-class parts of the learning objective — not setup work performed by a human outside the loop.
Commodity motors and electronics, digitally fabricated fluidic and structural parts. Precision metal actuator stacks only where physics genuinely demands them.
Four planes. Everything else on this page is a zoomed-in view of one box or one edge.
The interface between the bottom two planes is a product. Buy the substrate, put your own body, runtime and policy on top of it — or buy the whole machine.
For developers
Force, wherever you need it. BoidKit is the actuation and body-construction stack that makes a Limboid possible — sold as an architecture and a component ecosystem, so it can find its way into machines we would never have built ourselves.
Fluid power permits a separation that electromechanical joints cannot make: heavy energy-conversion hardware stays in the core, shared pressure is routed through the body, and lightweight distal actuators create force near the joint that needs it.
This does not make hydraulics automatically cheap or efficient. Seals, leakage, pressure losses, contamination, thermal management, valve precision, pump efficiency and fatigue all remain real engineering constraints. The claim is architectural: shared pressure makes a different cost and mass allocation possible.
Linear contractile fluid actuators, bundled in parallel. Instead of sizing one actuator for the full joint load, an axis uses several fibres per antagonistic side — so each fibre carries a fraction of the aggregate tensile requirement before load-sharing, transient and fatigue factors.
One 4/3 valve meters pressure and return between the two sides of an antagonistic joint. Progressive port geometry gives fine flow resolution off the null and large flow area at full rotation — fine and coarse behaviour from one cheap valve instead of duplicated metering hardware.
Shared pressure and return galleries, valve seats, sensor ports and mounting collapsed into one digitally fabricated unibody. Parameterised from the morphology description rather than redrawn each revision.
Two pump cassettes merge through check valves into a high-pressure accumulator, behind a pressure guard with relief, sensing and a normally-open dump path that bounds stored energy.
Water or water/glycol rather than petroleum hydraulic oil. Lower mess in human environments, easier self-fabrication of fluid components, and far safer early proof and burst testing.
Deterministic local control so main compute never micromanages valve PWM. Owns encoder sampling, valve zero and calibration maps, setpoint interpolation, limiting, fault detection and graceful behaviour when upstream packets are late.
The fluid network doubles as a proprioceptive medium. A root transducer injects a small pressure chirp into the same lines that carry power; each branch modulates its own reflection, and round-trip delay identifies which one answered.
Γi = ( Zm,i/Ap,i2 − Z0,i ) / ( Zm,i/Ap,i2 + Z0,i ) τi = 2Li / ci
Unresolved: dispersion, branch reflections, valve-state coupling, flow and pump noise, identifiability in dense manifolds, calibration drift. This is a research track, not a shipping feature.
Four electric tendon drives per hand: two opposing finger groups, thumb flexion, and thumb opposition that transitions between pinch and power grasp. Distal joints are coupled and compliant, so the hand conforms without commanding every phalanx.
Make the geometrically expensive parts, buy the commoditised ones. Manifolds, headers, muscle components, shells and tendon routing come off accessible digital-fabrication equipment. Cells, BLDC motors, bearings, MCUs, cameras and safety-rated electrical parts are bought.
The goal is not ideological avoidance of metal. It is to stop precision cylinders and custom geartrains from dominating the marginal cost of every joint.
BoidKit was conceived as a modular, plug-and-play actuation platform, and it stays one even as our own robots get more vertically integrated. If you already have a stack, use ours from the outside.
Whole machines
Two topologies on one substrate. Both are designed around human affordances — doors, counters, shelves, tools, handoffs — without pretending to be human. The form is synthetic, feature-minimal and visibly a machine.
Quadrupedal base · humanoid upper body
Four ground contacts hold a large support region while the upper body works. That buys stability, payload and a far lower control burden than exquisitely tuned dynamic bipedal balance — which is why it is the platform we are building first. “Centaur” describes topology only: four legs plus a human-compatible upper body. There is no engineering reason for a tail, fur or hooves.
Bipedal humanoid
The humanoid is where this ends up: the narrowest footprint, the widest set of human spaces, the fewest environments that need modifying. It is also the harder control problem, and it is honest to say we get there through the centaur rather than instead of it. Same substrate, same runtime, same hands — different topology in the morphology graph.
Mains stays offboard during charging; the chassis receives isolated low-voltage DC through an interlocked inlet, BMS, service disconnect, precharge and main contactor.
Starting bands, not constants. Hydraulic bandwidth, valve geometry, sensor latency, flexible modes and contact regime set the real stable rates.
An independent safety supervisor holds authority over energetic subsystems regardless of main-compute health. Every major fault converges on three hardware actions:
Backed by passive layers: relief valve, fuses, check valves, load-hold geometry on major axes, mechanical joint stops, pressure-rated containment.
A display, not a mechanism. Arbitrary gaze and expression, high-bandwidth status, readable accessibility modes — and an explicit, honest indication of whether the machine is running autonomously or under someone's direct control. Cheaper and more semantically honest than an actuated pseudo-face.
Structural skeleton, then the actuator and fibre layer, then a coarse elastic support mesh, then a conformal smoothing layer, then a waterproof cleanable membrane. Form emerges from the real frame and routing rather than decorative armour — and stops well short of simulated human anatomy.
Self-fabricated fluid-power parts get disciplined gates before they touch a robot:
fabricate dimensional inspection hydrostatic proof behind shielding leak & decay cycle / fatigue calibration verified safe-stop release
Home fabrication does not mean skipping pressure-vessel engineering.
Embodied autonomy
In software, the hard parts are already discretised into click, call, run. In robotics, “pick up the object” expands into camera calibration, segmentation, pose uncertainty, two coordinate transforms, trajectory generation, backlash, contact detection, friction, grasp closure, load verification, and recovery from slip. So we treat measurement construction as part of the task rather than as invisible infrastructure.
Camera intrinsics, extrinsics, tool-centre transforms, encoder zeros, valve deadband, pressure curves, contact bias and timing offsets are represented explicitly, with uncertainty — not assumed.
calibration:
head_camera_extrinsic:
estimate: …
covariance: …
observed_at: …
left_gripper_tcp:
estimate: …
covariance: …
valve_zero:
zone_2_axis_4: …
time_sync_error_ms: …This lets the planner ask a question most stacks cannot: am I limited by uncertainty about the world, or about myself?
Observations, motor commands, calibration changes, hypotheses, tool calls, safety events and model updates become a typed causal history the machine can audit itself against.
Event
event_id · timestamp · stream_id
stream_type ∈ {sensory, motor, tool,
calibration, thought, judge,
train, safety}
parent_event_ids · causal_tags
uncertainty · calibration_state
world_state_delta · belief_delta
self_model_deltaThe property we want: “the cup is in the sink” can point back to the action, the contact events, the pose estimates and the verification frames that support it.
The learning objective rewards bringing the internal transition model into alignment with the world's actual transition structure through intervention.
Not a finished objective function. A direction: reward the system for making reality more predictable through grounded intervention.
Nothing learned should hard-code one monolithic vector of joints and sensors. A Limboid describes itself as a typed morphology graph — rigid bodies, joints, muscle bundles, valves, pressure zones, tactile surfaces, cameras, IMUs, tendons and compute nodes, connected by kinematic, hydraulic, electrical, sensing, control and contact edges.
No pre-order theatre
Every claim on this page carries one of four labels. We would rather be checkable than impressive.
This is not a name attached to a newly imagined robot. BoidKit went through roughly eighteen months of intensive R&D from early 2023 into mid-2024, then continued as sustained side-project development.
That thesis has not changed. What changed is scope: from an actuation kit to a complete machine — body, manipulation, autonomy, fabrication.
A technical dependency order, not a set of dates.
Older public material from this program quoted very low actuator and whole-robot manufacturing figures. We are not repeating them. They were exploratory estimates, not measurements, and they will not reappear here until they are rebuilt from a current line-item BOM and a validated prototype.
A sub-$1,000 machine at manufacturing scale is our economic target. It is not a demonstrated cost, and we will say so every time we say the number.
These are the metrics we intend to publish instead, as prototypes mature:
What muscle geometry maximises lifetime-adjusted force per dollar, rather than peak force alone?
Can water-hydraulic seals and valves stay low-leakage over useful cycle counts with mostly polymer components?
What pressure minimises total system mass once pump, line, valve, actuator and structural scaling are all counted?
Can a progressive rotary valve deliver both fine low-flow control and high peak flow from cheap servo torque?
Which axes can be passive or compliant without materially hurting manipulation?
Are four drives per hand enough for the everyday object distribution?
Can static joint loads be held mechanically, so standing still costs no hydraulic power?
What joint-loop bandwidth is achievable with compliant fibres and low-cost valves?
Can individual backscatter branch responses be separated under active pump flow?
Which physical skills survive fleet replay without catastrophic sim-to-real brittleness?
What real duty cycle dominates energy use: locomotion, static support, manipulation, or compute?
How much pressure sensing is actually required if position and contact data are good?
Test plots, failures, CAD revisions, BOM changes and the occasional working joint — sent when there is something real to show, and not otherwise. No launch countdowns.