# protoAdapt — initial maze-navigation comparison

Experiment date: September 24, 2026. Public research note prepared September 25, 2026.

## Question and apparatus

Can an inherited neural controller navigate unfamiliar mazes, and how do final-generation winners from two training habitats perform across habitats?

This is a local development experiment with abstract propulsion and steering. It is separate from the MuJoCo/MJX walking studies and from protoAdapt's experimental muscle-body mode. The simulated body is a rigid tetrahedral marker with variable body parameters and a reactive neural controller. The model uses simplified environmental and movement rules, not validated robotics physics.

## Training

- Two runs: meadow and wetlands.
- Same initial population, experiment seed 42.
- 48 candidates; 30 evaluated generations; three shared randomized worlds per generation.
- Each trial lasts 24 simulated seconds.
- Maze layout and conditions are regenerated between generations. Generation scores therefore do not form a directly comparable all-time performance curve.
- Selection uses 0.75 × mean trial score + 0.25 × worst trial score.
- Tournament selection, crossover, Gaussian mutation, elites, and random immigrants produce the next generation. Elites are evaluated again on the new worlds.
- The genome has 220 values: body/rhythm parameters, oscillator phases, and an 18-input, 8-hidden-unit, 6-output neural controller. Navigation uses two outputs for propulsion and steering; the remaining outputs are neutral in this mode.
- Sensors include five local obstacle rays, goal bearing and straight-line distance, speed, terrain, friction, wind, water, rhythm, body size, and accumulated effort. No maze map or shortest path enters the controller.
- The evaluator can access shortest-path distance. Reward combines net and peak route progress, goal arrival, limited exploration reward, and effort/collision penalties. Weights do not learn within a lifetime.

## Evaluation

The last completed generation's winner from each run was tested for 24 seconds on seeds 3000000019, 3200000087, and 3400000099 in each of meadow, highlands, wetlands, and ice conditions. These fixed seeds were held out from the training runs.

The comparison baseline is eight randomly initialized founder genomes, tested on those same three worlds in each habitat. Score is the mean composite reward, not a percentage. The reports record the number of goals reached and mean progress for each evolved controller; they do not include per-trial observations or founder goal counts.

The source JSON field `generation: 29` uses zero-based indexing and refers to the thirtieth evaluated generation. Both winners reached 9 of 12 test goals. The website table is generated from the retained JSON reports, with scores rounded to two decimal places.

## Interpretation and limits

Both evolved controllers scored above the random-founder mean in these tests. The comparison does not establish habitat specialization, statistical significance, or optimal behavior. It uses one training seed and only three fixed maze seeds per habitat. The two training runs began with the same genomes and are a paired diagnostic comparison, not independent replicates.

Further study needs multiple independent training seeds, equal evaluation budgets, fresh test environments, and uncertainty estimates. When fixed test scores guide further development, treat those environments as a validation set and reserve new cases for final evaluation.

## Retained public records

- [Meadow-trained benchmark](https://protobytes.art/research/records/meadow-42.benchmark.json)
- [Wetland-trained benchmark](https://protobytes.art/research/records/wetlands-42.benchmark.json)
- [Experiment notebook](https://protobytes.art/research/experiments/)

These downloads are aggregate reports. Complete checkpoints and experiment code remain in the local development archive; this release is not a complete reproduction package.