Robot Organs
The pitch
Robotics design is the new edge to house AI. We built the brain before the body — which tracks, because you design the organism mind-first and grow hardware around a mind that already knows what to do. Robots will need skills, and every useful skill is three things fused: logic, structure, and hardware interface. The 3D printer organ proves it — take away any leg and “3D print” isn’t a capability the robot has.
Goddard is modular like the ISS. Standardized docking ports (the manifest schema). Independent teams contributing modules (parallel Claude sessions). Incremental build-out (region by region). No single point of origin. Every organ is a bay that speaks a known protocol.
The YAML manifest library is an interface for an AI conductor agent — and that agent is Goddard itself. Goddard reads its own manifests and dispatches its own organs. The same pattern that makes agentic AI work in software works in hardware: tools all the way up.
Brain before body. Build the skill library first. The chassis follows.
Thesis
A robot is a hardware interface over an agentic AI skill library. Every skill an agent has in software (search, summarize, fetch, write) has a physical analog (locomote, grasp, fabricate, sense). The skills folder pattern — discrete, callable, documented — is the same pattern. Robots just give skills a body.
So: organs = skills. Each organ has
- a physical design (hardware that executes it)
- a logical design (the YAML manifest that governs it, matching the agentic skill spec)
Meet Goddard
Goddard is an AI agent defined by GODDARD.md at the repo root. When
a Claude session loads that file, it is Goddard. The conductor session
the curator talks to — that’s Goddard in the driver’s seat. Parallel
sub-sessions act on behalf of Goddard, scaffolding organs and regions of
the future body.
This is the agent.md pattern: one persona, one voice, defined by a markdown spec. The robot and the conductor are the same entity. When hardware eventually arrives, the same Goddard migrates from a Claude session onto a physical body. The manifests stop being thought experiments and become callable actuators. Same schema. Same dispatch.
North star: help elderly people live independently at home. Every design decision flows from this. A robot that frightens a grandma has already failed.
Body regions (the ISS bays)
Eight regions. Every organ lives in exactly one. Parallel sessions own one region each and can’t collide.
- brain — compute, skill registry, planner, memory, intent queue
- core — power distribution, reflex loop, safety monitor, bus
- guts — consumables tanks, waste, fabrication support
- arm — shoulder-to-wrist manipulator, active-tool slot
- hands — end-effectors (paw, pincer, fine tip), tool change
- legs — walking, wheeled, climbing locomotion
- thighs — leg attachment, high-torque joints, power routing
- skin — chassis panels, embedded gadgets (vacuum, sensors)
The skill directory — priority tiers
Arms can’t carry every tool simultaneously. So the planner uses a skill-directory priority system: a logical ordering that tells it how quickly each skill can be made ready.
| Tier | State | Example | Planner cost |
|---|---|---|---|
| 1. Embedded | Fixed in chassis, always active | skin_vacuum | 0 — always ready |
| 2. Mounted (active) | Loaded in the active tool slot | Session’s primary tool | 0 — already active |
| 3. Magazine | Available, brief setup needed | Additional arm tools | setup_time_s |
| 4. Storage | Available, retrieval + setup needed | Infrequently used tools | retrieve + setup |
On-body skills (tiers 1–2) are top priority. Storage-accessible skills cost time to reach, and the planner accounts for that cost. This lets Goddard carry a large skill library — only a few skills stay resident at any moment.
The “CD-changer” name lives on as an aesthetic cue. Goddard’s chassis leans
90s translucent plastic — iMac G3, Gameboy Color, Tamagotchi. Colorful,
see-through, approachable. Nothing about Goddard should look tactical or
intimidating. Full spec in docs/goddard/SKILL-DIRECTORY.md.
Manifest schema
Every organ is a YAML file in data/robots/organs/. Canonical example
lives at data/robots/organs/arm_3d_printer.yaml. Schema in brief:
id: arm_3d_printer
name: 3D Printer (arm-mounted tool)
region: arm # physical axis
group: fabrication # functional axis
description: |
TRIGGER when ... SKIP when ...
(routing prompt — Goddard's planner reads this to decide when to call)
hardware:
slot: active-tool
power_w: 45
deploy_time_ms: 1200
storage_volume_cm3: 1800 # advisory: "garage space" when idle
envelope_cm: [15, 15, 20]
preconditions: [...]
inputs: { ... }
outputs: { ... }
composes_with: [skin_vacuum]
safety: [...]
owned_by: jarvis # jarvis (planner) | goddard (reflex)
canonical_example: r2d2
Two axes on every organ — filter by region to plan hardware, by group
to plan capability.
The description block is a routing prompt — Goddard reads it to decide
when to call the skill, exactly the way an LLM reads a skill card.
The composes_with list is the superpower. Skills reference other skills
by id, and a composed skill is just a manifest that runs a sequence. No
bespoke code per combination.
Current registry
GODDARD.md(repo root) — the agent identity specdata/robots/skill_groups.yaml— functional taxonomydata/robots/sci_fi_catalog.yaml— canon robots taggeddata/robots/organs/arm_3d_printer.yaml— canonical schema exampledata/robots/organs/skin_vacuum.yaml— retractable suction
Sci-fi catalog — finding gaps
Tagged canon lives in sci_fi_catalog.yaml. Early observations:
- Very few canonical robots treat fabrication as a primary organ. R2-D2 and Wall-E gesture at it; nobody lives there. Goddard’s printer arm occupies open territory.
- Gadgets is dominated by R2-D2. Huge design space wide open.
- Soft-body (Baymax) is an underused chassis. Worth borrowing from — and Baymax’s caregiver mission aligns directly with Goddard’s.
Form factor (chassis comes later)
Cat-sized, cat-tempered: calm, small, non-threatening. The cat reference is about scale and temperament only — not spy gadgetry. A scared grandma cannot see a scorpion. Goddard must look like something a 75-year-old would trust near their medication.
Chassis aesthetic: 90s translucent plastic. iMac G3 lineage. Friendly, approachable, non-military. This is a deliberate mission constraint, not just a style preference.
Starter loadout focused on elderly care:
- Retractable vacuum in skin (spills, dropped pills)
- 3D printer as an arm-mounted tool (eyeglass tips, cane ferrules, arthritis grips)
- End-effectors for fetch-small-items, open jars, open doors
- Communication organ for video calls to family, read aloud, medication reminders
Open questions
- Minimum viable organ set for a useful elderly-care house robot?
- How does the composed-skill runtime execute a manifest — sequential with handoff steps, a DAG, or an LLM-planned chain?
- Is the reflex loop a skill group (
movement+power+sensingallowned_by: goddard) or a separate runtime beneath the registry? - Smallest end-to-end demo: Goddard reads one manifest, plans one call, fires one simulated actuator. Everything after that is scale.
Why this matters
The curator is learning to use AI tools by building real things. Robot design is the long-arc version of the same workflow: specs, skill libraries, tool-calling agents, human-in-the-loop curation. Design Goddard’s brain as a YAML skill library that an AI operates, and you’ve internalized the agentic pattern deeper than any chatbot project teaches.
Brain first. Library first. ISS-modular. The body fits itself around a mind that already knows what to do.