Rule 01
Vendor-agnostic from day one.
Reachy Mini is the wedge, not the commitment. Skills, brains and the policy engine live above an abstraction so that adding a new robot is a new bridge, not a rewrite.
§ Robotics · A playground
A body for the brain. Carefully.
Kapllan Robotics is the small, exploratory wing of the lab where we ask what changes when our agents stop being chat windows and start being objects in a room. This is not a product roadmap. It is a playground, a notebook, and a record of what we learn as we learn it.
● Status: Reachy Mini reserved 2026-05-04First robot · Fiaelli (Wireless)
§ I — The first body
Reachy Mini — wireless. Named Fiaelli, by my daughter.
The wedge is small on purpose. Reachy Mini is a head with two antennas, a camera, a microphone, and a speaker — and the quiet expressivity that comes from those four things together. We are starting here because the failure modes are legible and the cost of being wrong is low.
Pollen Robotics built the body. We bring the brains — voice, reasoning, memory — and the policy layer that decides what a small robot in a home is permitted to do.
Hardware · Pollen Reachy Mini, WirelessBrain · gpt-oss-20b voice + Ofelia reasoningCompute · CM4 onboard, LLM off-robotReachable · Public from v1, via authenticated tunnel
§ II — Open questions
Five things we are trying to find out.
- Q · 01
What changes when an agent has a body?
We have brains — reasoning, voice, vision. We don't yet know what it feels like when one of them can also nod, look at you, or turn toward the door. The first thing we want to learn is what we don't already.
- Q · 02
How small can a useful presence be?
Reachy Mini is not a humanoid. It is a head, two antennas, a speaker, and a microphone. A modest body. We are curious whether that is enough — for company, for a small assistant, for a child's first hello.
- Q · 03
When should an agent initiate?
A chat window is a polite servant — it speaks only when spoken to. A robot in a room cannot pretend to be invisible. We will need to decide what it is allowed to volunteer, and what it is required to keep to itself.
- Q · 04
What does emotion-mapping feel like in person?
Hand-tuned baselines, LLM-emitted tags, sentiment fallback — three layers stacked. On screen, this kind of thing is design polish. In the same room, it is the difference between a friendly object and an unsettling one.
- Q · 05
What is the right safety posture for a body in the home?
Robots are higher-stakes than file edits. We are bringing over the deny-by-default policy engine we built for our coding agent and rethinking it for motion, audio, and voice identity. Humility is a feature.
§ III — A journey
Five layers, each one teaching us the next.
We are not promising features by date. We are publishing the order in which we plan to learn them, and a guess at how long each one will take.
- v1On arrival + 1–2 weeks
Voice companion
What we hope to learn · What it is to talk to a thing in the room and have it listen back. Tool-calling — lights, rooms, audio — earns the right to be embodied.
- v2+2–3 weeks
Second-brain bridge
What we hope to learn · Whether asking out loud — “what did I learn about X?” — feels different from typing the same question into NxNotes.
- v3+3–4 weeks
Outreach
What we hope to learn · When an agent should initiate. A finished training run, a milestone hit, a pull request that needs eyes. The robot becomes an outbound channel, not just an inbound one.
- v4Month 2
Voice ID
What we hope to learn · What changes when the robot knows who is talking. Speaker diarization, an identity database, and the manners that follow from both.
- v5Month 3+
Family bridge
What we hope to learn · The emotional north star. A child says “pappa” into one body and the message arrives in another. Wake-word, identity verification, secure peer routing — all built so that this last step is layered, not rewritten.
§ IV — Working rules
How we are building this. Plainly.
Rule 02
No new intelligence — only plumbing.
The brains already exist. Voice, reasoning, second-brain RAG, boardroom deliberation. Robotics is the layer that lets those brains nod, listen and gesture. We are not training new models for it.
Rule 03
Safety as a first-class module.
Deny-by-default motion envelopes, blast-radius classification, an audit log. Borrowed wholesale from our coding agent and adapted for a body. Robots are higher stakes than file edits.
§ V — Field notes
We will write here as we learn here.
Robotics work surfaces in the research catalog as field notes. The voice agent essay was the first; expect short, dated entries when something teaches us something — including the failures, especially the failures.
Read the field notes →