Assumptions, and what is derived

Assumptions in this lab are layered and not fixed. There is a common core (this page) that holds across experiments, plus per-experiment specifics that add to or override it — because the right assumptions depend on the experimental setup. Each experiment page states its own deltas; this page is the shared reference.

Every result traces to one of three things: (A) physics the engine integrates, (B) a calculation our control code performs, or (C) a modelling idealization.

Common assumptions (unless an experiment says otherwise)

Planar rigid-body world. The agent is one rigid body with three DOF (slide_x, slide_y, yaw) and viscous damping; MuJoCo integrates the dynamics (RK4, dt = 5 ms). No out-of-plane motion (z, roll, pitch); no ground friction (frictionless glide, resisted only by joint damping).
Whiskers are range sensors, not bristles. Each is a zero-thickness, rigid, straight rangefinder ray — unbounded but capped at MAX_RAY (4 m). It reports distance to the nearest surface, not contact or bending along a length. No flex, no tactile surface.
The world is a walled arena (plus objects) at body height; rays hit walls and objects.
Reproducible. All RNGs are seeded; runs are byte-reproducible.
Data is logged every timestep, whether the agent is moving or not.

A. Simulated by MuJoCo (derived)

Rigid-body planar dynamics (pose/velocity from integrated forces+damping); rangefinder distances (raycasts against geoms); proprioception (a velocimeter + gyro giving true body-frame linear velocity and yaw rate).

B. Computed by our control code (not physics)

The drive wrench (net force+torque from the located drive zones' positions); the dead-reckoned pose (integrating proprioception); the world model (hit points placed from estimated pose ⊕ body-schema whisker pose ⊕ measured distance); the scores (coverage/precision as point-based nearest-neighbour fractions, ε = 6 cm, against ground-truth surface samples).

C. Idealizations (common)

Movement is a computed wrench on a rigid body — no simulated limbs, muscle–tendon units, or ground-contact thrust.
Whiskers are distal rays, not tactile bristles (no flex, no contact).
Proprioception is idealized (true velocity, optionally + noise); the SMN "balance/co-activation generates data at rest" (Register 1) is not modelled.
The occupancy map is a monotonic accumulator by default; an optional decay rate turns it into a living snapshot — see Living snapshot.
The agent has no biological state by default. The Basal coupling experiment adds an energy reserve, BAP gating by energy, and food in the world — the constructive assumptions for why the agent moves at all.
The HAP does not by default read the world model. The Map-guided foraging experiment adds a food-memory living-snapshot layer (populated by consumption events) and a steering term that biases the turn toward live remembered locations — turning the map's decay rate into a direct survival pressure.
The visual experiments use one undifferentiated camera + one flat 8 × 8 modulator as a deliberate floor. In SMN, perceptual resolution is a function of CAZ density and internal capacities, not of the raw transducer — and since unmodulated input is dropped, raising the camera's pixel count alone would change nothing. The next visual experiments raise CAZ density on the eye itself and add known asymmetries between eyes; see P0-visual.
The eye has its own CAZ pair from P1-visual onward. The forward model predicts the next frame from the sum of head and eye yaw displacements, using a per-column sec²(α) correction so the warp tracks rotation faithfully across the whole FOV. Saccades are generated by a minimal random-fixation controller.
The modulator's per-patch floor calibration is the architecture's first-order "scene memory" — see P1-visual / static-world. It detects any departure from the calibration distribution (motion, novel appearance, novel persistent content with edges the floor wasn't calibrated against); it does not habituate on its own (a slowly-updating floor would be a natural successor).

Per-experiment assumptions (these vary)

aspect	P0 reafference	P0-visual	P1-visual	P1 world model	P2 self-localized	sweep
body	single CAZ "head"	same single CAZ "head"	head + eye CAZ pair	mobile, 5 whiskers	mobile, body schema	varies (3/5/9 whiskers; track width)
locomotion	none (head only)	none (head only)	none (head only)	central `xfrc` thrust	located opponent drive zones	+ `±BAP` toggle
steering	opponent yaw (motors)	opponent yaw (motors)	head + eye opponent yaw	opponent yaw (motors)	differential located drive	+ routing: flat / hierarchical
transducer	one rangefinder whisker	camera at the head's pivot (128 × 128, fov 90°); flat 8 × 8 patch grid	camera at the eye's pivot, smoothed (~σ 1.0); flat 8 × 8 patch grid	5 whisker rangefinders	5-whisker fan	varies (whisker count)
forward model	binned (heading → range)	analytical frame-warp from `ω_z · Δt · focal_px`	angle-correct per-column shift `Δθ · focal_px · sec²(α)` from `Δθ_head + Δθ_eye`	n/a	n/a	n/a
world dynamics	static + scheduled exafference object	static + oscillating exafference object	static + oscillating exafference object	static arena	static arena	static arena
internal model	learned forward model + residual	floor-gated 8 × 8 patch residual	floor-gated 8 × 8 patch residual	occupancy accumulator	occupancy accumulator	+ `±HAP` toggle

The P1 → P2 shift is the clearest example that assumptions track the setup: P1 builds the map from the simulator's true pose; P2 uses only the agent's own proprioceptive self-localization. Same task, different assumption — and it changes what the experiment can claim.

D. How the figures are computed

The result maps (p0/p1/p2/sweep) are matplotlib plots of computed arrays (hit cloud, trajectories, ground-truth outlines, metric text). No physics render.
agent_geometry.png is a matplotlib schematic from MouseSchema.
scene_render.png is a real MuJoCo (EGL) rasterization with the rangefinder rays drawn by the visualizer.

Tonic / co-activation data at rest (Register 1).
~~World-model decay rate (a living snapshot in dynamic equilibrium).~~ — ✅ available now via OccupancyMap(decay=…); see Living snapshot. 2b. ~~A biological state that ties cognition to viability (why move?).~~ — ✅ available now via energy + BAP gating + food in Basal coupling.
Physical, flexible whiskers that bend and sense contact along their length.
Actuated locomotion (real muscle/tendon + ground contact).
Out-of-plane DOFs (z, roll, pitch); friction; uneven terrain.
Sensor models: rangefinder noise/dropout; proprioceptive bias/drift.

F. What is faithful to SMN

Opponency at the zones (in-place rotation needs antagonists); reafference self/world from an efference-keyed forward model (P0); located zones + explicit body schema; a world model built in relation to body geometry from the agent's own self-localization; BAP and HAP both necessary to build a world model (ablations collapse coverage).