Roadmap

Rolling versioning: 0.MINOR.PATCH, incremented as work lands. A major version bump (1.0.0, 2.0.0, ...) is a deliberate decision made by a human — never automated.

We work one release at a time. Don't start tasks from a later release until the current one is closed out, unless explicitly reprioritized.

Closed releases are one line each in CHANGELOG.md. This file is open work only.

Status

Layer	Fact
Implemented	20 Hz loop: sense → physiology → drives → memory → action → actuate → learn. Learned sensor→actuator reflexes trained by drive reduction. Named-channel body. SoftAP dashboard. Life-state JSON. Station firmware as a separate state machine.
Tested	Host-native tests for bit packing, age saturation, rate-limit clamp, LEDC cap, and organism behavior (energy dynamics, contingency bias, reflex learning rule) against a fake body. Not real hardware.
Simulated	The unmodified engine in a modelled room with a station (`src/sim`): learned reflexes lift mean lifespan from 1.26 h to 1.86 h (hand-wired bound 2.46 h). Learns to reach and stay on the dock; not yet to leave when full, and nothing accumulates across lives.
Hardware tested	Nothing. No board has run this.
Designed / planned	Body declared in a JSON file with per-input ranges, stepper and mic-envelope channel kinds (example build); approach-and-dock on hardware; extra boards.
Hypothesis	That drives + decaying memory + learned reflexes + noise will look alive on real hardware, including charging as an attractor. Unverified.

v0.10.0 — Simulator and learned reflexes (current)

Native simulator driving the unmodified engine in a modelled room with a station, ablation conditions, a no-engine baseline and a hand-wired upper bound (src/sim)
Learned sensor→actuator reflexes trained by drive reduction, with an eligibility trace
Physiology fixes the simulator exposed: hunger band, hunger→activity, energy double count, sensory adaptation and startle habituation, curiosity out of the reward, convex drive, correlated exploration noise
Example build: 7 inputs, 5 outputs on one ESP32 DevKit
Save reflex weights in life-state (today they're lost on reboot)
Leave when full: the learned organism stays docked at 87% charge
Learning that accumulates across lives instead of peaking in the first one
Reflex weights on the dashboard

v0.11.0 — Body from a file

body.json on LittleFS, editable from the dashboard, applied live without reflash
Per-input range (battery 9.0–12.6 V → 0..1)
envelope input for an analog mic (background sampling, RMS)
stepper output (velocity mode, shared enable, disabled at rest)
ESP-NOW station beacon for ~10 Hz RSSI (the simulator can measure what the 5 s scan costs)

v0.9.1 — Hardware validation (blocked, no hardware in this environment)

End-to-end test: emergent approach-and-dock behavior over repeated sessions

Split out of v0.9.0 rather than left blocking it, since everything else in that release was done and independently verifiable, and this one task isn't a firmware task at all — it's "build the physical robot and station, run them together, and watch what happens over repeated sessions."

Blocked, honestly: needs a real robot, a real station, and repeated physical trials — there is no way to verify this from firmware code review or PlatformIO builds, and this environment has no hardware to run it on. Every prior release has carried a "not yet run on real hardware" caveat; this is the first task that can't even be partially exercised without hardware, since it's fundamentally about physical behavior over time, not firmware correctness. Left unchecked rather than claimed done. This is also, not coincidentally, exactly the kind of validation v1.0.0 is gated on below — this milestone stays open until that hardware exists.

v1.0.0 — First stable "alive" release (human-gated)

All of the above validated on real hardware
Experimental program from the paper (synth-behavior.md §15) run at least once: homeostasis quality, exploration structure, history dependence, perceived aliveness
Ablation switches implemented (disable stochasticity, freeze contingency memory, clear spatial memory, flatten drives, decouple battery from h_energy)
Human sign-off to cut 1.0.0

Backlog / later

Multi-agent / multi-robot stigmergic experiments (scent, RF, light tags)
Mechanical voice / prosody experiments
Additional board support as hardware becomes available
Formal comparison against HRL baselines