The object's true outline is never rendered in embodied mode. A jointed arm sweeps a probe around it at a fixed angular rate; at each discrete step the probe's tip touches the surface and reports one distance reading, exactly like a fingertip tracing an object with the eyes closed. Each touch drops a point at the contact location and the growing polyline through those points is the agent's internal model of the shape — a sensorimotor contingency: knowledge encoded not as a picture but as "if I turn my arm by Δθ, contact moves by this much."
touch(θ) = raycast(origin, dir(θ)) + noise
model += point(touch(θ))
coverage = touches / samplesPerRevolution
- Exploration speed — angular rate of the arm sweep; faster sweeps cover the contour sooner but each touch still contributes only one point, so the model still needs a full revolution.
- Shape — the object being explored: circle, square, triangle or a concave five-point star, chosen without ever being shown.
- Embodied exploration vs instant vision — toggling to instant vision renders the true mesh immediately, the way an unobstructed glance would: recognition is instantaneous because perception did not have to be built action by action.
- Model accuracy — how closely the touch-built point cloud tracks the true boundary; each reading carries small sensorimotor noise, so the reconstructed contour is close but never pixel-perfect, unlike direct vision.
This contrasts two accounts of perception: the disembodied/computational view, where a shape is recognized by matching a static image against stored representations, and the embodied/enactive view (Varela, Thompson & Rosch), where knowing a shape *is* the accumulated sensorimotor law linking motor commands to sensory consequences — perception constructed through action, not read off a picture.