maxgfr/skycmd-trajectories

skycmd game trajectories (models vs oracle)

Complete episodes of the four skycmd games (Flappy Drone, Rescue Run, Obstacle Dodge, a Doom-like shooter) played by four small skycmd decision models and by the exact oracle: at every decision, the state, the model's probabilities, the oracle's soft target, the reward and the events. Generated by our Rust game cores and models: no third-party data.

Part of skycmd, tiny decision models trained from scratch in Rust that run in the browser. Written by scripts/gen-trajectories.sh (sky-arena <model> --trajectories, seed 2026) at commit bd5a64a1. Licences of every set: DATA_LICENSES.md.

One config per policy, one row per decision, in play order:

field	meaning
model, game, episode, step	who plays, which game, episode index, decision index in the episode
tick	simulator ticks before the decision (flappy decides every tick, doom every 4)
state	the compact text state the policy reads (the same as in the browser)
question, options	the action question and its options, rendered as text
probs	the policy's probability per option
oracle	the oracle's soft target per option, computed from the full simulator
action, oracle_action	the action played (argmax of probs) and the oracle's argmax
abort_probs, abort, oracle_abort	dodge only: the abort question, answered by both
reward, events	reward summed until the next decision, and what happened (passed, crashed, rescued, kill, hurt, ...)
vetoed	the coded safety reflex replaced the action
last	last decision of the episode

The episodes_* configs give one row per episode: decisions, return (every tick's reward), truncated (flappy stops alive at 50 pipes), score, max_score, success, crashed, aborted, safe_abort, vetoes, steps. Flappy has no mission to complete: its episodes stop alive at 50 pipes (truncated), so its completion is 0 and its score is the pipes passed.

What it is for. Every state a small model reaches by itself, with the expert's answer next to its own: imitation learning and DAgger (oracle is the correction), offline reinforcement learning (rewards and outcomes from five policies of very different quality, including complete failures: s1-pico and s2-nano never learned doom), calibration (confidence against correctness along an episode), and the study of compounding errors.

Per policy and game

policy	game	episodes	decisions	agreement with the oracle	mean score	mean return	completion	crash
oracle	flappy	20	107,520	1.000	50	103.76	0.00	0.00
oracle	rescue	200	8,704	1.000	3	35.65	1.00	0.00
oracle	dodge	100	33,223	1.000	9.23	13.79	0.78	0.00
oracle	doom	100	11,672	1.000	2.96	13.03	1.00	0.00
s1-pico	flappy	20	107,520	0.946	50	103.76	0.00	0.00
s1-pico	rescue	200	9,313	0.907	2.995	35.24	0.99	0.00
s1-pico	dodge	100	34,043	0.845	9.2	13.71	0.77	0.00
s1-pico	doom	100	119,835	0.000	0.02	-1.30	0.00	0.12
s2-nano	flappy	20	103,942	0.943	48.3	100.22	0.00	0.05
s2-nano	rescue	200	8,760	0.961	3	35.62	1.00	0.00
s2-nano	dodge	100	37,561	0.765	9	13.31	0.73	0.00
s2-nano	doom	100	125,314	0.000	0	-0.65	0.00	0.06
s2-nano-doom	flappy	20	107,520	0.946	50	103.76	0.00	0.00
s2-nano-doom	rescue	200	8,860	0.960	3	35.52	0.99	0.00
s2-nano-doom	dodge	100	34,560	0.814	9.01	13.33	0.75	0.00
s2-nano-doom	doom	100	11,800	0.917	2.96	13.04	1.00	0.00
s3-pico	flappy	20	107,520	0.950	50	103.76	0.00	0.00
s3-pico	rescue	200	8,896	0.963	3	35.55	1.00	0.00
s3-pico	dodge	100	35,058	0.830	9.12	13.58	0.76	0.00
s3-pico	doom	100	11,791	0.911	2.96	13.03	1.00	0.00

Sources and licences

Each config is one source and keeps its source's licence (share-alike sources stay share-alike). Attribution goes to the upstream authors; the conversion is ours.

config	upstream	licence	how it is used	rows
episodes_oracle	ours: oracle	apache-2.0	one row per episode: return, score, outcome	420
episodes_s1-pico	ours: s1-pico	apache-2.0	one row per episode: return, score, outcome	420
episodes_s2-nano	ours: s2-nano	apache-2.0	one row per episode: return, score, outcome	420
episodes_s2-nano-doom	ours: s2-nano-doom	apache-2.0	one row per episode: return, score, outcome	420
episodes_s3-pico	ours: s3-pico	apache-2.0	one row per episode: return, score, outcome	420
oracle	ours: the games' exact oracle, which reads the full simulator (expert demonstrations)	apache-2.0	every decision of 20 flappy, 200 rescue, 100 dodge and 100 doom episodes	161,119
s1-pico	ours: skycmd s1-pico (ModernBERT-style, trained on flappy, rescue and dodge; never saw doom)	apache-2.0	every decision of 20 flappy, 200 rescue, 100 dodge and 100 doom episodes	270,711
s2-nano	ours: skycmd s2-nano (trained on flappy, rescue and dodge; never saw doom)	apache-2.0	every decision of 20 flappy, 200 rescue, 100 dodge and 100 doom episodes	275,577
s2-nano-doom	ours: skycmd s2-nano fine-tuned on doom with DAgger	apache-2.0	every decision of 20 flappy, 200 rescue, 100 dodge and 100 doom episodes	162,740
s3-pico	ours: skycmd s3-pico (the four games)	apache-2.0	every decision of 20 flappy, 200 rescue, 100 dodge and 100 doom episodes	163,265

Files

file	rows
data/episodes_oracle.parquet	420
data/episodes_s1-pico.parquet	420
data/episodes_s2-nano.parquet	420
data/episodes_s2-nano-doom.parquet	420
data/episodes_s3-pico.parquet	420
data/oracle.parquet	161,119
data/s1-pico.parquet	270,711
data/s2-nano.parquet	275,577
data/s2-nano-doom.parquet	162,740
data/s3-pico.parquet	163,265
files/summary.json	—

from datasets import load_dataset
ds = load_dataset("maxgfr/skycmd-trajectories", "episodes_oracle")

skycmd wrote no row with a language model: the targets come from the upstream labels (some upstream sets carry their own teacher soft labels, see their cards), the games' oracles or our generators.