Adapters

Adapters are the contract between your signals and the brain. An input adapter writes onto an afferent port; an output adapter reads an efferent port or a decoder on the intrinsic population. Standard adapters ship with the engine; custom adapters are yours.

Standard input adapters

Adapter	Port	What it does
`io.Screen`	visual	Coarse full-screen map plus a foveal window around a point, resampled onto the hexagonal photoreceptor grid.
`io.Camera`, `io.Frames`	visual	Video frames onto the compound eye, with optional motion pre-emphasis.
`io.Microphone`, `io.Audio`	mechanosensory	Audio onto the Johnston's organ afferents (antennae), 16 kHz mono.
`io.Embedding`	olfactory	Any vector, up to 2,282 dimensions, from a co-processor: tokens, instructions, seed points.
`io.BodyState`	ascending	Proprioception: cursor position, controller state, joint encoders.
`io.IMU`	mechanosensory	Accelerometer and gyro as mechanosensory input.

Standard output adapters

Adapter	Port	What it does
`io.Mouse`	descending	dx, dy and three button states at 100 Hz.
`io.Keyboard`	descending	Key down and key up events with timing.
`io.Gamepad`	descending	Two sticks, triggers and buttons.
`io.Motors`	descending, motor	Torque or position targets per joint, up to 1 kHz.
`io.Readout`	central, optic	A small trained decoder on the intrinsic population: boxes, events, segments.

Custom adapters

A custom adapter is a small module with an encode or decode method and a declared port. The engine checks the port size at load time and trains the adapter together with the weights.

Python

import flycore, torch

class Lidar(flycore.InputAdapter):
    port = "visual"                       # 11,390 afferents
    def encode(self, scan: torch.Tensor) -> torch.Tensor:
        # scan: (N, 2) ranges and angles -> (11390,) activations
        return self.hex_project(scan)

brain = flycore.load("flylabs/servo", adapters={"lidar": Lidar()})

Co-processors

A co-processor is a classical network that runs before or beside the brain and feeds it through an adapter, almost always the embedding port. Intent (a vision-language model), the speech recogniser in Dictate and the CNN classifier in Gaze are co-processors. They are ordinary PyTorch modules; you can replace them.

Python

intent = flycore.load("flylabs/intent")
cursor = flycore.load("flylabs/cursor", coprocessors=[intent])
cursor.do("click Save")