You can deploy this model and start earning money today!

Karajan42/
dream-robot-gauntlet-policies

Name: dream-robot-gauntlet-policies
Availability: InStock

robotics

world-models

spiking-neural-networks

evolution

sim2real

cc-by-nc-sa-4.0

Views

No views yet

Model Card for Dream Robot Gauntlet: Policies

17 CMA-ES-evolved Leaky-Integrate-and-Fire (LIF) spiking controller genomes for a differential-drive robot, evolved entirely inside a neural DOOM world model, plus the controller math, two dream-eyes vision probes, and the Webots verification environment used to test the genomes against real rigid-body physics.

Model Details

Model Description

Each genome is a CMA-ES-evolved parameter vector for a small LIF spiking network that maps proximity-sensor inputs to differential-drive motor commands (see controller/lif4.py for the exact architecture and decode). Two sensing regimes were evolved: oracle sensing, where fitness is scored against the simulator's exact state, and dream-pixel sensing, where a fitted linear probe reads virtual left/right proximity values directly off the world model's own rendered frames, and fitness is scored from that probe-reconstructed signal.

Developed by: Alakazam
Model type: Evolved spiking neural controller (LIF) + two linear "dream-eyes" probes (pixel-to-proximity)
License: CC BY-NC-SA 4.0, inherited from the DOOM world-model checkpoints this was evolved against
Training method: CMA-ES (evolution strategy), not gradient-based RL

Dependencies (not included in this repository)

This release does not include the world-model checkpoints these genomes were evolved and probed against (~280MB denoiser.onnx each). Fetch them separately:

Karajan42/doom-engine-arena: deathmatch-trained checkpoint (deathmatch/denoiser.onnx). The dream-eyes probe fitted against this checkpoint reaches r_L = 0.903, r_R = 0.922 (n = 400 pairs), with a missed-hot-state rate of 0.09 (left) and 0.03 (right).
Karajan42/doom-dungeon-55: the "v55" checkpoint (denoiser.onnx). The probe fitted against this checkpoint reaches r_L = 0.785, r_R = 0.869 (n = 320 pairs), with a missed-hot-state rate of 0.0 on both channels.

Folder	Contents
`champions/`	17 evolved genomes. `champion_doom_oracle.json` (fitness 1.642, oracle exact-state sensing) and `champion_dream_eyes.json` (fitness 2.357, dream-pixel sensing, world `dream-eyes-margin`), plus 15 `oracle_genomes_.json` files, one per world/sensing-regime variant tried (`doom-final`, `doom-balanced`, `doom-h25`, `dream-eyes-latch-native` at fitness 2.377, the highest reached across the whole sweep, and others). Each file stores the CMA-ES genome vector and its fitness score (labeled `doom_F`/`dream_F` in the `champion_.json` pair and `phi` in the `oracle_genomes_.json` files; both names carry the same Floreano-lineage metric, used at different pipeline stages); the `oracle_genomes_.json` files additionally record a target world name and a champion flag.
`probes/`	`probe_W.npy` and `probe_W_v55.npy`: the fitted linear probe weight matrices (one per world-model checkpoint above) that convert dream-rendered pixels into virtual left/right proximity readings. Paired training data for these fits is in the sibling dataset repo's `probes/` folder.
`controller/`	`lif4.py`: the LIF controller math, including the no-stop decode and the turn-commitment latch that fixed a stalling bug (see below). `s5_train.py`: the training driver these controllers were evolved with.
`env/`	Six-file doom-gym Python environment (`doom_env.py`, `doom_env_v55.py`, `evolve_doom.py`, `evolve_dream_eyes.py`, `probe_collect.py`, `validate_extractor.py`) wrapping the world-model checkpoints above in a gym-style step/reset interface for both CMA-ES evolution and probe fitting.
`webots/`	Three `.wbt` world files (`gate3_arena`, `gate3_slalom`, `gate3_field`) plus `gate3_lif.py`, the Webots controller wrapper that runs the same LIF math from `controller/lif4.py` against real rigid-body physics; this is the independent substrate the champions above are verified against.
`media/`	Driving clips from both substrates (`dream_drive.gif`, `dream_drive_v55.gif`, `dream_eyes_drive.gif`, `webots_dream_eyes_final.mp4`), a contact sheet (`v55_contact_sheet.png`), and five report-quality figures (`fig_alignment_ladder.png`, `fig_champion_scorecard.png`, `fig_probe_scatter.png`, `fig_jitter_beforeafter.png`, `fig_hero_grid.png`) built from `measurements.json`. These five supersede the earlier `alignment_ladder.png` and `webots_trajectories.png` (kept for reference); the `fig_*` versions carry the corrected titles and captions and are what the technical report and blog post cite.

Evolution fitness ladder

Five sensing/reward configurations were evolved with CMA-ES over roughly 3,500 local episodes total, at $0 marginal compute cost (local hardware, no cloud spend). Fitness reached: oracle exact-state sensing, 1.642 (champion_doom_oracle.json); dream-pixel sensing at 1.907 (balanced), 2.344 (v1), 2.357 (margin, champion_dream_eyes.json), and 2.377 (latch-native, oracle_genomes_de2.json, the highest fitness reached in the sweep). This fitness is the CMA-ES evolution objective (Φ, Floreano-lineage), not the Webots physics-verification pass/fail described next; the two are separate measurements and are not interchangeable.

Physics verification

24 of these genomes were run through Webots (full per-run data in the sibling dataset repo's oracle_runs/), 20 episodes in an open arena plus 20 in a slalom course per genome. None cleared the full joint gate (zero contacts, minimum clearance above 2cm, at least 95% of steps at or above 5cm clearance, and at least 2m path traveled, required jointly in both worlds).

Seven genomes recorded zero contacts across all 40 episodes in both worlds, but they split into two distinct families rather than one uniform pattern.

Parkers (five oracle runs, three genome files): oracle_genomes_final.json (Φ = 2.169, world doom-final), oracle_genomes_h25.json (Φ = 1.642, world doom-h25, evaluated at its default control rate and again at 350ms), and oracle_genomes_h8.json (Φ = 1.838, world doom-h8, evaluated at 250ms and 350ms) hold position near their start: mean path traveled is 0.06m–0.22m across arena and slalom for all five of these runs.

Drivers (two genomes, this release's flagships): oracle_genomes_de2.json (Φ = 2.377, world dream-eyes-latch-native) drives with mean paths of 2.02m (arena) and 2.36m (slalom); champion_dream_eyes.json (Φ = 2.357, world dream-eyes-margin) drives with mean paths of 1.63m (arena) and 2.31m (slalom). Neither driver's gate miss is immobility. The margin-final champion's arena path bar is missed on the mean (1.63m against the 2.0m bar), dragged down by a worst-episode path low of 6.4cm, not by typically parking; the latch-native champion's worst-episode path low is 1.29m in the arena and 1.93m in the slalom, both still substantial fractions of the 2.0m bar. Both drivers, like all seven zero-contact genomes, miss the 2cm minimum-clearance floor in the slalom course (clearances of 0.89cm–1.3cm across all seven); two of the seven (the latch-native champion and the 350ms evaluation of oracle_genomes_h25.json) also miss that floor in the arena (1.55cm and 1.94cm respectively).

Zero contacts repeats across all seven genomes, parkers and drivers alike; a genome that both clears contacts and drives the full course at full clearance margin is the open item this release does not close.

Turn-commitment latch

A latch added to the LIF decode in controller/lif4.py removed a stalling failure mode. Before the fix, 64% and 53% of a 180-tick episode were spent in frozen/non-committal turn ticks across two traces; after the fix, 0 frozen ticks in both post-fix traces (dataset repo's logs/jitter*.jsonl).

Uses

Direct Use

Load a champion genome and controller/lif4.py to reproduce the evolved controller's behavior, either in the doom-gym environment (env/, requires one of the world-model checkpoint dependencies above) or in Webots (webots/ .wbt worlds plus gate3_lif.py; no world-model dependency needed, since Webots is the independent verification substrate).

Out-of-Scope Use

No genome in this release has cleared the full joint physics-verification gate. Treat every champion here as a research artifact demonstrating zero-contact navigation under a partial gate, not a certified obstacle-avoidance controller.

Bias, Risks, and Limitations

No genome passes the full joint verification gate (contacts + clearance + coverage + path-length, jointly, in both Webots worlds); see Physics verification above.
The fitness ladder (oracle sensing at 1.642 vs dream-pixel sensing up to 2.377) reflects the CMA-ES objective's own reward shaping under each sensing regime measured here; it is not a general claim that dream-pixel sensing outperforms exact-state sensing on the physics-verification gate.
Non-commercial license, inherited from the DOOM world-model checkpoint lineage.

Citation

Technical report not yet published; cite this repository directly in the interim.

bibtex

@misc{hernandez2026dreamrobotpolicies,
  title  = {Dream Robot Gauntlet: Policies},
  author = {Hern\'andez, Hugo and Alakazam},
  year   = {2026},
  note   = {Private pre-publication model release}
}

Dataset repo (raw runs, selection history, probe pairs): Karajan42/dream-robot-gauntlet-data
World-model dependencies: Karajan42/doom-engine-arena, Karajan42/doom-dungeon-55

Model Card Authors

Alakazam (alakazam.gg)

Karajan42/dream-robot-gauntlet-policies