Appendix B
The machines
Axis counts, hands, sensing, the power tree, how control is layered, and what happens in hardware when something goes wrong. The centaur is under construction; the humanoid inherits nearly all of it.
1 · Why a centaur first
Engineering target
Four legs carry the machine and one humanoid torso does the work. Stability, payload and a far lower control burden come from below; reach, manipulation and compatibility with rooms built for people come from above. The term centaur covers the topology and stops there. Nothing about it argues for a tail, fur, hooves or equine proportion — each of those costs mass. Feet are broad contact structures, shaped by traction, compliance, sensing and whether they can be made at all.
For useful physical labour, locomotion and manipulation must coexist. The quadrupedal base trades some humanlike appearance for substantially easier static stability, higher payload, more support polygons, and much less dependence on exquisitely controlled dynamic balance. Whole-body control at low speed is far more forgiving when four ground contacts can hold a large support region while the upper body works.
2 · Body map
| Group | Count | Notes |
|---|---|---|
| Legs | 4 × ~3 hydraulic | Hip pitch and roll, plus knee. |
| Torso | yaw + pitch | Carries compute and the fluid core. |
| Neck | yaw + pitch | Carries the head, cameras and face display. |
| Arms | 2 × ~7 hydraulic | Wrist axes included in the count. |
| Hands | 2 × 4 electric | Tendon drives, not hydraulic. |
| Distal joints | passive / compliant | Where an actively controlled axis does not justify its cost. |
| Total | ~30 hydraulic + 8 electric | The bipedal Limboid lands at the same totals with 2 × ~6 leg axes. |
3 · Hands
Engineering target
The machine needs hands that are actually useful, not merely hand-like. The minimum useful capability is reliable power grasp, hook grasp, pinch and basic fine manipulation. Full independent actuation of every phalanx is deliberately avoided in the first architecture because it explodes cost, wiring, sensing and control complexity.
One concrete design uses four electric tendon drives per hand:
- Opposing finger group A flexion
- Opposing finger group B flexion
- Thumb flexion and pinch force
- Thumb opposition, transitioning between pinch and power grasp
Distal finger joints are coupled and compliant. Tactile pads and tendon elasticity let the shape conform without independently commanding every joint. Feedback comes from fingertip force and contact pads, servo position or tendon displacement, optional flex sensing, and wrist pose with an arm force estimate. This is not anatomically complete; it is much closer to the actual task requirement, which is grasping objects robustly at low cost.
4 · Form and face
Long-term vision
The intended upper body is smooth and synthetic rather than military or industrial. Form should emerge from the actual frame, joint volumes, soft tissue equivalents, routing and protective skin — not from decorative armour plates. The anticipated construction is layered: structural skeleton and joints, distributed actuator and fibre layer, a coarse elastic support mesh that contains and organises the soft mechanical layer, a finer conformal layer that smooths local geometry, and a waterproof cleanable outer membrane.
The result may read as uninterrupted and form-fitting, but without sex characteristics or simulated skin anatomy. The machine should look like a coherent nonhuman organism-shaped tool rather than a person in armour.
Rather than mechanically imitating a face, the head uses a display as its main expressive surface. That gives arbitrary gaze and expression rendering, high-bandwidth status communication, readable telepresence cues, accessibility modes, role indication, and an explicit signal of whether the robot is autonomous or under direct control. It is both cheaper and more semantically honest than an actuated pseudo-face.
5 · Sensing
Exact sensors are modular. The architecture assumes combinations of head and torso RGB cameras, depth or stereo where economics justify it, IMU, microphones, joint encoders, pressure and temperature sensing, contact and tactile sensors, motor current and power telemetry, optional external calibration cameras or fiducials, and radio signal observations.
Camera intrinsics, camera-to-body extrinsics, tool-centre-point transforms, encoder zero, valve deadband, pressure and flow curves, contact sensor bias and timing offsets are represented explicitly, with uncertainty. See Appendix C.
6 · Electrical architecture
Concrete engineering design, subject to revision
| Chemistry / configuration | 16s LiFePO₄ |
| Nominal voltage | 51.2 V |
| Full charge | 58.4 V |
| Nominal capacity | 32 Ah |
| Stored energy | ≈ 1.64 kWh |
All propulsion-bus components need voltage margin above maximum charged voltage. LiFePO₄ is attractive for cycle life and thermal behaviour, though pack mass and power capability still need evaluating against the final duty cycle.
Mains voltage stays offboard during normal charging; the chassis receives isolated low-voltage DC.
7 · Real-time control
High-level intelligence is separated from deterministic control. High-level compute is allowed to be probabilistic and occasionally slow. The plant interface is not.
| Inner joint / valve loop | ~100–200 Hz |
| Whole-body setpoints | ~20–100 Hz |
| Visual perception | camera rate, e.g. ~30 Hz |
| Semantic planning | variable, asynchronous |
| Cloud reasoning / training | never assumed to meet a hard deadline |
These are starting bands, not universal constants. Hydraulic bandwidth, valve geometry, sensor latency, flexible modes, contact regime and body scale determine the actual stable loop rates.
Whole-body control
Task-level objectives are converted into feasible body commands while respecting:
- Contact constraints and support-polygon balance margin
- Joint and tendon limits
- Hydraulic pressure and flow budget
- Self-collision
- Payload estimate and manipulation forces
- Body compliance
- Current and thermal limits
8 · Safety
Engineering target
Safety is a separate control authority. An independent supervisor holds power over energetic subsystems regardless of main-compute health. All major energetic faults converge on three hardware actions:
- Inhibit pump torque
- Open the hydraulic dump path
- Open the main battery contactor
Fault inputs include emergency stop, pressure excursion, leak detection, overtemperature, motor overcurrent, watchdog timeout, bus failure, implausible encoder or pressure state, and body-impact conditions.
Passive layers sit underneath: relief valve, fuses, check valves, load-hold geometry on major axes, mechanical joint stops, pressure-rated containment, and low-energy-safe default valve states where possible.
Physical safety dominates task completion. Remote instructions and learned policies cannot bypass the independent safety layer, and stopping never depends on a network round trip.
9 · What a task actually costs
Consider: bring me the black backpack from the other room. This does not map to a single learned action token. It expands into:
- Task interpretation — which backpack, probable location, destination reference for “me”, permission to enter, conflict with current payload or tool state.
- Epistemic check — if location confidence is low, select an information-gathering action rather than committing to a grasp trajectory.
- Locomotion — body path under obstacle map, leg reach and contact constraints, body envelope, battery margin.
- Visual verification — identify the backpack and distinguish it from similar objects.
- Manipulation planning — graspable region, pose uncertainty, likely mass, required wrist pose, one hand or two.
- Contact action — reach and grasp under local control, monitoring tactile, contact and tendon response.
- Grasp verification — visual motion consistency, hand force, arm load, object-relative pose.
- Carry — whole-body control updated for shifted centre of mass and restricted arm workspace.
- Deliver and verify — place or hand off; verify release and final object state.
- Event update — store successful and failed subtraces for later retrieval and training.
Physical competence is a composition of calibrated perception, self-modelling, control, intervention and verification — not semantic instruction following.