HABITAT 001 / METHODS & LIMITS

A small, inspectable
behavioral system.

The homepage contains five simulated agents named Spring, Summer, Autumn, Winter, and Night in a circular arena with five stationary obstacles and a ramp onto the largest platform. They select destinations, turn gradually, brake near obstructions, and give one another space. Each agent has individual curiosity and personal-space weights.

What comes from the Blender work

The browser adaptation follows the structure of ASimulation’s Blender motor: a desired route, a capped steering bias, a forward clearance estimate, acceleration and braking, and bounded turning. The personal-space bias is capped at 0.75 relative to the goal direction, following the Blender concurrency correction.

Spatial familiarity is inspired by the prototype’s decaying exposure memory. Here it uses coarse three-unit cells, a 45-second exponential decay constant, and a maximum of 400 cells per agent. Candidate destinations favor lower familiarity as curiosity increases. This is a simplified adaptation, not a line-for-line port.

What is different here

The browser uses analytic circles and a ground plane instead of evaluated Blender meshes, BVHs, and a mesh path planner. Local tangential obstacle steering replaces global path planning. The browser advances in fixed 0.02-second steps; the Blender motor uses 0.05 seconds. It includes an in-place recovery turn when blocked, unlike the Blender motor’s strict moving turn radius.

Movement updates sequentially, so agent order can affect encounters. Temporary deadlocks and local detours are possible. The world has privileged access to other agents’ positions; it does not infer them from camera images. First-person view is an observation camera, not an implemented visual-perception model.

Observation-only controls

Visitors can select an agent, switch between overview, follow, and first-person views, and change camera rotation and elevation. There are no public controls for stepping, resetting, pausing, or changing agent drivers. The application exposes no global simulation-control API. The simulation is local to each visitor’s browser, not a shared persistent world.

Dialogue and contact

Names appear in short greetings when a pair enters a 2.8-unit proximity band and goodbyes when it separates beyond 3.7 units. These different thresholds prevent boundary jitter from repeatedly triggering dialogue. Pair events have 14-second cooldowns, and each speaker has a six-second cooldown. Wall remarks use an 18–30-second cooldown.

“Ow” requires a body contact event at an incoming speed above 0.65 units per second, with a ten-second contact cooldown. A collision guard prevents penetration. Proximity and wall remarks never count as hard contact. Contact speech can be suppressed by a speaker’s cooldown, and successful avoidance means it may not occur during a normal visit. These lines are authored feedback, not feelings or language understanding.

Ramp, platform, and jumping

One agent at a time can approach the ramp, climb onto the large platform, wander on top, and jump back to the ground. A visit targets 10–30 seconds on the platform, weighted by curiosity and a random component. The landing area is reserved; an agent waits at the edge if it is not clear. Follow and first-person cameras track its height.

This is an authored movement sequence with a curved jump trajectory, not a general rigid-body physics solver or learned jumping skill. Other obstacles remain non-climbable. Random lines about Flatland, enlightenment, personal growth, and the shadow self mark ascent and landing.

Who is speaking?

Spring is green, Summer red, Autumn blue, Winter white, and Night black. Speech bubbles retain the speaker name, use a matching outline, and connect to the moving agent with a pointer. A screen-space layout separates visible bubbles and keeps them inside the habitat, including on narrow screens.

Short conversations

Nearby ground agents can stop for about 12.5 seconds, face each other, greet one another by name, then exchange a randomly selected question and reply. Topics include dimensions, meaning, boundaries, and psychology. Each agent has a 45-second conversation cooldown. These are authored jokes and exchanges, not generated language, LDRM output, or evidence of spiritual or psychological experience.

Surface interest and recognition

Agents evaluate the ramp using the same geometric surface description that draws it. A ground-connected incline counts as travelable when its slope is at most 35 degrees and its width exceeds the body diameter by 0.2 units. Flat ground, vertical walls, narrow strips, and floating surfaces do not qualify as ramps. Interest rises smoothly with proximity over a 12-unit range, weighted by curiosity; interested agents compete for one available ramp visit.

This is explicit geometric classification, not learned understanding or visual perception. The classifier accepts rotated incline descriptions; the current habitat still has one authored ramp-to-platform route. It is not a general mesh-navigation system.

Gestures, memory, and spectators

Speech can trigger a small nod or head tilt, and playful replies can trigger a bounce. Agents remember conversation partners and witnessed jumps for 90 simulated seconds, then forget them. A nearby available agent can briefly watch a jump and offer a reaction. Cooldowns and availability checks keep these events from interrupting every conversation. All dialogue remains authored.

What this does not establish

These are authored rules with short-lived state. There is no reinforcement-learning optimizer, neural training, LLM, evolutionary reproduction, or demonstrated subjective experience. Apparent personality reflects different parameters and histories. This demonstration is not evidence of consciousness or general intelligence.

Research context

The local project’s Embodied Geometric Agent Ecology for Blender report proposes a broader architecture involving geometric perception, memory, specialist agents, and developmental experiments. The curve-projection memory research is a separate proposed substrate; it is not running in this browser. Broader affective, social, and meaning-related dimensions remain research questions.

Composable steering has an established precedent in Craig Reynolds’s steering-behavior work ↗.

Bounded validation

The underlying planar steering regression (without excursions or conversation pauses) used twenty deterministic seeds, which were run for five simulated minutes each. Every agent traveled at least 11.28 world units in every 30-second window. Minimum peer center distance was 1.12 units; minimum obstacle surface clearance was 0.586 units, against a 0.56-unit body radius. The longest near-stationary spell was 21.30 seconds. Separate fixtures checked hard contact, recovery, proximity hysteresis, dialogue cooldowns, and deterministic replay. These tests cover this fixed arena, not arbitrary maps or general crowd navigation.

The combined ramp and conversation system was separately tested across eight seeds for ten simulated minutes each. Every run completed repeated climbs and landings and at least one conversation. Checks covered finite coordinates, ground height after landing, bounded per-step displacement, three-dimensional peer spacing, and measured platform dwell times within 10–30 seconds. These tests cover the authored route in this arena; they do not establish general terrain traversal.

Privacy & dependencies

The simulation runs in your browser. This site sets no cookies, uses no analytics, and does not save your simulation state to a server. The hosting provider processes normal web requests. Following an external link takes you to that service’s policies.

3D rendering uses locally hosted Three.js, under its MIT license. Fonts are system fonts.