# Research comparison for DustFly / mesocortAI

Reviewed 2026-09-13. This is a source review, not an independent reproduction of the referenced projects.

## Assessment

Using anatomical wiring as a constrained computational substrate, with engineered sensors and trained input/output interfaces, is a legitimate experimental design. A showcase can publish our implementation, exact information boundaries, frozen evaluations, intervention results, and synchronized replays. Neither a real connectome nor similarity to another project establishes biological fidelity or a performance advantage over a conventional controller.

Our V6–V8 question is concrete: can a fresh image/pose/goal interface, coupled through the full retained fly graph to a PFL3-only motor interface, reach new continuous goals on a known 3D Dust2 replica? The selected weights and final paired outcomes are recorded separately. The browser replays computed trials; the local GPU lab computes new goals.

## Primary sources

| Source | What the source reports | Relationship to this experiment |
|---|---|---|
| [Lappalainen et al., Nature 2024](https://www.nature.com/articles/s41586-024-07939-3), [official flyvis documentation](https://turagalab.github.io/flyvis/) | A connectome-constrained visual-system model is optimized for optic flow, then assessed against neural responses. | Strong evidence that task optimization of anatomical models is a research method. Different circuit, task, dynamics and validation; not validation of DustFly. |
| [FlyDoom](https://github.com/eganeganegan/flydoom) | Describes visual encoders, sparse anatomical recurrence, fixed/trainable/readout-only modes, PPO, and matched graph/controller controls. Initial subgraphs are smaller than our retained graph. | Closest methodological parallel to learned interfaces around a recurrent connectome. We reviewed its documentation, not its benchmark reproducibility. |
| [fly-craftax](https://github.com/liuzihe02/fly-craftax) | Documents PPO training of a linear descending-neuron readout, held-out controls, and a scrubbable activity viewer. | Supports the relevance of learning motor interfaces rather than assuming anatomy already supplies useful game controls. |
| [DoomFly](https://github.com/nftechie/doomfly) | Reports 166,700 retained cells, 25,582,938 directed edges, engineered sensory/motor assignments and plasticity on 4,184 KC→MBON11 connections. Its reviewed README reports failed visual, conditioning and survival gates for its candidate V6. | Same retained graph scale and useful disclosure practice. It does not establish successful learned survival, and its V6 is unrelated to our version numbering. |
| [Awesome Fly](https://github.com/cobanov/awesome-fly) | Curated index spanning full graphs, small subsets, games, body models and research code. | Discovery aid, not certification. Project inclusion is not scientific validation. |
| [DAgger, Ross et al.](https://arxiv.org/abs/1011.0686) | Dataset aggregation addresses distribution shift by obtaining expert labels at states visited by the learner. | Basis for V7. Teacher access is restricted to training cases. |
| [PPO, Schulman et al.](https://arxiv.org/abs/1707.06347) | Policy optimization with a clipped objective. | Basis for V5 and V8; our environment, reward, hierarchy and anatomy integration are engineering choices. |

Other projects are methodological references. Their performance numbers or claimed topology benefits are not presented as DustFly results; our measurements come from our own runs. The flybody geometry has its separately included Apache 2.0 attribution.

## Claims we can make

- The retained anatomical graph really participates in the simulated control loop; its input and motor interfaces are explicit.
- V3 changes a restricted set of existing internal weights. V6/V7 train a fresh image/pose/goal adapter. V8 trains sensory and motor policies jointly with PPO around fixed original anatomy.
- Prepared replay activity is sampled from simulation, with older reconstructions separately identified.
- New held-out results apply to this map, training procedure, fixed checkpoints, physics, information access and evaluation protocol.

## Claims requiring further experiments

- Biological wiring outperforms a matched, equally trained alternative. Frozen-policy lesions do not answer this; retrained matched controls and multiple seeds are needed.
- Visual perception matches a living fly. Our depth camera, coordinate inputs, sensory mapping, signs and rate dynamics are engineered approximations.
- RL trains all internal synapses. V8 does not do that: its original anatomical weights are fixed.
- Any point is guaranteed reachable by the learned controller, or that it generalizes to another map. A finite known-map test cannot establish either.
- The system plays full Counter-Strike. This is goal-reaching navigation in our 3D Dust2 replica.

## Český závěr

Přístup má smysl zveřejnit jako dohledatelný navigační experiment: anatomická síť je součástí řízení a rozhraní se učí různými způsoby. Ostatní projekty ukazují metodické souvislosti, nikoli důkaz našich výsledků. Přesná tvrzení musí vycházet z našich zmrazených testů. Zatím netvrdíme biologickou věrnost, převahu mušší topologie ani zaručenou cestu k libovolnému cíli.