Tracking

What must the controller do?

Tracking moves an observer around a circle to follow a stimulus. The world stores observer angle theta and stimulus angle phi in radians. The stimulus moves one degree per world tick by default and reverses direction every 720 ticks.

Two eyes have default offsets of ±30 degrees. Each eye samples 31 directions from −60 to +60 degrees in four-degree increments, giving 62 receptor channels. Each receptor has a four-degree central response plateau and a Gaussian fall-off outside it. sensory_gain scales this signal.

The two effectors move the observer by movement_amp * (left - right) degrees per tick; the default amplitude is 10. There is one neural frame per world tick.

from brainlesslab import CompositionSpec, NodeSpec, TaskSpec, profile, resolve

target = resolve(profile(CompositionSpec(
    node=NodeSpec("falandays"),
    task=TaskSpec("tracking", options={"randomize_start": False}),
    count=32,
))).targets[0]
print(target.horizon, target.n_inputs, target.n_effectors)

Expect 2000 62 2. Disabling random starts sets observer angle π/2, stimulus angle zero and direction +1 unless explicitly overridden. The ordinary default randomises the initial angles and direction.

Read the native outcome

track_score is the mean cosine of angular error during the scored interval. It ranges from −1 to 1: perfect alignment is 1, perpendicular alignment is zero and opposite alignment is −1. The minimum scored horizon is 2,000 ticks. Warm-up may precede scoring and still changes the world and adapting reservoir.

The random-action reference samples a uniform signed movement command each world tick. Matching empirical calibration is required for a null-adjusted score. The upper bound 1 is a mathematical bound, not a measured oracle.

A short diagnostic trajectory can show motion and interface behaviour. It cannot establish ordinary task performance. Continue with a paired plasticity comparison and calibration.