Karajan42/dream-robot-gauntlet-data

Dataset Card for Dream Robot Gauntlet: Data

Raw episode logs, physics-verification runs, probe-fitting pairs, and evolution selection history from a campaign that evolves a spiking robot controller entirely inside a neural DOOM world model, then verifies the result in Webots, a rigid-body physics simulator that shares no code with the world model.

Dataset Summary

The controller is a Leaky-Integrate-and-Fire (LIF) spiking network, evolved with CMA-ES against two different sensing regimes: exact simulator state ("oracle" sensing) and virtual proximity values read directly off the world model's own rendered frames through a fitted linear probe ("dream-pixel" sensing, the "dream-eyes" mechanism). Every evolved genome is then run through Webots, an independent physics substrate the dream never saw during training, to check whether behavior learned inside the dream transfers to real rigid-body dynamics. This repository is the data half of that campaign; the evolved genomes, controller code, and probe weights are in the sibling model repository (linked below).

Contents

Folder	Contents
oracle_runs/	24 Webots physics-verification runs, one directory per evolved genome. Each holds arena.jsonl (20 episodes, open arena), slalom.jsonl (20 episodes, slalom course), gate3_arena.log + gate3_slalom.log (simulator run logs), and verdict.json (the joint pass/fail gate result).
probes/	probe_data.npz and probe_data_v55.npz: paired (dream-frame, ground-truth-proximity) samples used to fit the two dream-eyes linear probes. extractor_pairs.npy: paired feature-extractor activations used upstream of the probe fit. Fitted probe weights live in the model repo, not here.
maps/	replica_arena.json, replica_slalom.json, corridor_map.json: the exact Webots physical layouts that the doom-gym environment (env/ in the model repo) replays as an oracle for exact-state fitness scoring.
selection/	candidates.json plus 18 sel_ev_*.json files: per-generation, per-candidate CMA-ES selection scores from the evolution runs.
logs/	55 run logs (chain, oracle_*_run, localgym_*, movie_run*, doomev_*, h8_oracles), 4 jitter*.jsonl traces, and traj.jsonl: day-by-day campaign runs, including the turn-commitment-latch fix (below).
lab_notebook/	RESEARCH_LOOP.md and REPORT-dream-robot.md: the running research log and internal report draft this card's numbers were pulled from.
measurements.json	The single source of every number in this card and its sibling model card. Built from the 24 verdict.json files plus the champion genome files; nothing here is transcribed from prose.

Headline numbers

Every number below is read directly from measurements.json.

Physics verification (oracle_runs/)

24 evolved genomes were each run through Webots for 20 episodes in an open arena and 20 in a slalom course (40 episodes per genome). The joint pass gate requires, per world: zero contacts, minimum clearance above 2cm, at least 95% of steps at or above 5cm clearance, and at least 2m of path traveled. Zero of the 24 genomes cleared the full joint gate.

Seven of the 24 recorded zero contacts across all 40 episodes in both worlds: oracle_de2, oracle_de3, oracle_final, oracle_h25, oracle_h25_350, oracle_h8_250, oracle_h8_350. These seven split into two distinct families, not one uniform pattern.

Parkers: oracle_final, oracle_h25, oracle_h25_350, oracle_h8_250, oracle_h8_350 (three underlying genomes: oracle_genomes_final.json, oracle_genomes_h25.json at two control rates, oracle_genomes_h8.json at two control rates). These hold position near their start: mean path traveled runs 0.06m–0.22m across arena and slalom for all five (verdict.json, path_mean_m). oracle_final is representative: 0 contacts across all 40 episodes, mean path 0.064m in the arena and 0.068m in the slalom.

Drivers: oracle_de2 (genome oracle_genomes_de2.json, Φ = 2.377, world dream-eyes-latch-native) and oracle_de3 (genome champion_dream_eyes.json in the model repo, Φ = 2.357, world dream-eyes-margin). These actually drive the course: mean paths of 2.02m (arena) / 2.36m (slalom) for oracle_de2, and 1.63m (arena) / 2.31m (slalom) for oracle_de3. Their gate misses are not immobility: oracle_de3's arena mean (1.63m) is dragged below the 2.0m bar by a worst-episode path low of 6.4cm, not by typically parking; oracle_de2'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.

In every one of the seven, the gate is still missed on the 95%-of-steps-at- 5cm-clearance requirement in both worlds, and on the 2m path-length minimum in at least one world. All seven, parkers and drivers alike, also miss the 2cm minimum-clearance margin in the slalom course despite zero registered contacts (slalom min_clear_m between 0.89cm and 1.3cm); two of the seven (oracle_de2, oracle_h25_350) miss that same margin in the arena too.

Zero contacts repeats across seven independent genomes, parkers and drivers alike, and is not a fluke of one run; a genome that both avoids every contact and drives the full course at full clearance margin is the item this release does not close.

Do not read "zero contacts" as "drives normally." Every zero-contact result in oracle_runs/ is scoped to one of the two families above: near-stationary parking, or driving with the movement gate still missed. The oracle's path gate is checked per episode, on path_min_m (every episode ≥ 2m), not on path_mean_m: oracle_de2 and oracle_de3 do post a path_mean_m at or above 2m in one or both worlds (see Drivers, above), but neither clears path_min_m in both worlds, so no zero-contact genome in this release clears the movement gate.

Turn-commitment-latch fix (logs/jitter*.jsonl)

Before a turn-commitment latch was added to the controller decode (see controller/lif4.py in the model repo), frozen ticks (stalled or oscillating turn decisions) accounted for 64% and 53% of a 180-tick episode across two traces (jitter.jsonl vs jitter2.jsonl). After the latch, frozen ticks measured 0 in both post-latch traces (jitter3.jsonl, jitter4.jsonl).

Provenance & caveats

• Collected July 18–19, 2026, on the local Alakazam doom-gym harness, a DIAMOND-style neural world model of DOOM (not the original game) running on local hardware.
• The two world-model checkpoints this data was generated against are not included (~280MB denoiser.onnx each). They are separate releases: Karajan42/doom-engine-arena (deathmatch/denoiser.onnx) and Karajan42/doom-dungeon-55 (denoiser.onnx, the "v55" checkpoint). The deathmatch vs v55 split in probes/ tracks which checkpoint each probe-fitting pair was collected against.
• oracle_runs/ is a verification set, not a training set: it is the output of running already-evolved genomes through Webots once each, 40 episodes per genome. The CMA-ES training-time rollouts themselves are not included; selection/ holds the per-candidate scoring from those runs instead.
• This campaign (the "doom" / Act II local port) is a separate experiment track from an earlier lingbot-based campaign (Act I). Results from the two should not be conflated when citing either.

Related

• Model repo (evolved genomes, controller code, probe weights): Karajan42/dream-robot-gauntlet-policies
• World-model dependencies: Karajan42/doom-engine-arena, Karajan42/doom-dungeon-55

License

CC BY-NC-SA 4.0, matching the DOOM world-model checkpoint lineage this data was generated against.

Citation

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

@misc{hernandez2026dreamrobotdata,
  title  = {Dream Robot Gauntlet: Data},
  author = {Hern\'andez, Hugo and Alakazam},
  year   = {2026},
  note   = {Private pre-publication dataset release}
}

Dataset Card Authors

Alakazam (alakazam.gg)