1. What you are seeing
The left panel is the Meta-Life world. The right panel shows the controller state: c is fast activation and m is slower memory.
Both fields are 8×8 and persist over time. The colored boxes are the learned spatial interface, drawn on c only:
- Red boxes I0..I3 are input ports. At the start of a macrostep, each input scalar xk overwrites its cell by a pointwise affine encoding operation: c[rowk,colk] ← wk·xk + bk.
- Blue boxes O0,O1 are output ports. After T microsteps, the decoder reads those c cells and applies pointwise affine maps: yℓ = wℓ·c[rowℓ,colℓ] + bℓ.
2. Meta-Life
The agent senses resources with three egocentric lidar rays:
- I0: left ray
- I1: center ray
- I2: right ray
- I3: reward feedback
3. Controller protocol
Each macrostep follows the order:
- sense lidar and reward
- encode inputs by affine overwrite into red ports of c
- run T CEM2D microsteps
- decode action from blue ports of c: O0 → steer, O1 → speed
- optional action noise, then clipping and movement
4. CEM2D microstep
σ = mean of the 8 neighboring c cells
q = a1 + a2·c + a3·σ + a4·m + a5·c·σ
c′ = c + α·(a0 + tanh(q) − λ·c)
g = sigmoid(gb + gk·q)
m′ = (1 − β·g)·m + β·g·c
a0..a5 define the shared local rule. α updates activation, β writes memory, and λ is fixed leak.