Droscape

A MaleCNS-derived steering model connected to a 2004Scape client

Droscape agent / 2004Scape
Measured connectivityloading
reading MaleCNS extractno substitute graph is shown
visual + ascending afferentsDNa01 / DNa02
Current actionDecision 000
Outputwaiting for client

No controller response yet.

DNa L−Rreward+0.000

The model path is target bearing → MaleCNS visual/ascending cells → 3,781 measured synapses with transmitter-informed modeled signs → bilateral DNa01/DNa02 activity → steering. Game-to-neuron transduction, sign interpretation, rate dynamics, adaptation and the final tile/action adapter are modeled—not measurements from a living fly.

Recent Controller Decisions

Each row is emitted by the running gateway controller. Dispatch proves the client received an action; reward comes from subsequent world-state changes.

stepoperationoutputreward
000waiting for clientcontroller offline

The character and controller state come from one server process. Visitors observe the same stream; opening or muting the player does not start another run.

Loaded Steering Circuit

The cells, sides, directed edges, synapse counts and transmitter predictions below come from the official Google Research / Janelia MaleCNS release. The extract is reproducible; the diagram geometry is only a layout.

The site could not load its connectome extract, so it is withholding the graph.

Provenance: official MaleCNS downloads. The extraction recipe pins every body ID and records hashes for the full 151,856,684-edge table, annotations and neurotransmitter predictions.

References

External material used for the dataset, biological interpretation and benchmark framing. Droscape is not affiliated with these projects or their authors.

What Enters the Network

RS-SDK exposes a target bearing. The bridge turns that bearing into left/right drive on named visual-projection and ascending cells. Activity then crosses measured edges into DNa01 and DNa02, with signs modeled from the transmitter predictions. The DNa left–right difference controls the next turn; bilateral activity controls stride.

Reward changes only the modeled gain at the RuneScape-to-afferent interface. It never rewrites the connectome weights. Chopping, dialog and inventory safety are declared game adapters, not claimed fly behaviors.

observation = {
  target:          —,
  heading:         —,
  desired_heading: —,
  left_drive:      —,
  right_drive:     —
}

What Would Count as Learning?

The topology comes from measured wiring, but “a fly learned RuneScape” would still be too strong: the neural dynamics and reward rule are modeled, and one character is not an experiment. A learning claim needs repeated trials and the controls below.

controllerpurposesuccess measure
MaleCNS topologytest measured wiringlogs per 100 trials
Shuffled topologycontrol for graph sizesame inputs and rewards
Matched RNNordinary ML baselinesame parameter budget

Task Design

Start with woodcutting

The current routine targets the nearest reachable tree. The DNa readout turns and advances the character one or two tiles at a time. Only when the character is adjacent does the bounded adapter issue “chop”; inventory safety and tutorial dialogs stay outside the neural claim.

What is biological

MaleCNS supplies the 16 traced cells, their bilateral identity, 60 directed edges, 3,781 synapses and transmitter predictions. DNa01/DNa02 are experimentally linked to low/high-gain steering. MaleCNS does not supply an executable mind: rate equations, game sensory encoding, reward adaptation and tile mapping remain explicit modeling choices.

What comes next

The client adapter emits observations and accepts bounded actions. The next research step is repeatable trial reset plus the shuffled-graph and matched-RNN controls.