README.md

---
license: cc-by-4.0
viewer: false
tags:
  - reinforcement-learning
  - offline-reinforcement-learning
  - unity-ml-agents
  - minari
  - causal-gpt-rl
---

# Causal GPT-RL — Unity ML-Agents trajectories

Recorded Unity ML-Agents trajectories packaged as
[Minari](https://minari.farama.org) datasets — **offline-RL datasets** spanning
continuous and discrete action spaces. Any offline-RL method can train on them; we
built them to develop **Causal GPT-RL**, our new approach that works across both
space types. Every environment
ships an `expert` tier; eight ship a full **quality ladder** (`expert` +
calibrated `medium`/`simple`) synthesized by degrading the stock policy — Gaussian
action-noise ranges for the continuous scenes, softmax-temperature ranges on the
policy logits for the discrete goal games.

Tiers are keyed to Minari-normalized skill between a random-policy `0.0` anchor
and the stock expert `1.0` — **targets `simple` 0.60 / `medium` 0.80 / `expert`
1.0** (envs land at simple 0.59–0.61, medium 0.80–0.82). The degradation is not a
single constant: it is drawn **per episode from a calibrated range** (the group/team
scenes — DungeonEscape, SoccerTwos — draw per match instead), so each tier's *mean*
hits its target while the episodes span a continuous band of skill instead of piling
at one point.

**Companion repos**
- Model-removed Unity builds + matching stock ONNX policies:
  [ccnets/causal-gpt-rl-unity-envs](https://huggingface.co/datasets/ccnets/causal-gpt-rl-unity-envs)
- Causal GPT-RL policy checkpoint (Pyramids):
  [ccnets/causal-gpt-rl-unity](https://huggingface.co/ccnets/causal-gpt-rl-unity)

## Data-quality notice — action labels (2026-07-30)

**All twelve datasets of the four discrete-action environments have been
re-recorded and replaced.** ML-Agents agents decide only every *k*-th physics
step and Unity repeats the last decision through the gap; the collector stored a
placeholder `0` for those in-between steps instead of the action Unity was
actually running. The action column therefore disagreed with the policy that
produced the trajectories — 72% of rows in the old PushBlock expert file.
**Observations, rewards, terminations and the trajectories themselves were never
affected**, so every published return, step-reward and tier calibration still
stands.

Share of rows that had held a placeholder action, measured per tier by replaying
the collector from the commit before the fix at each shipped band:

| env | affected rows | status |
|---|---:|---|
| `pushblock` | 47–72% | **all three tiers fixed** |
| `dungeon-escape` | 24–36% | **all three tiers fixed** |
| `soccer-twos` | ≈6% | **all three tiers fixed** |
| `pyramids` | 3–6% | **all three tiers fixed** |
| `crawler`, `walker`, `3dball-hard` | ≈1% | pending |
| `worm` | 0% | unaffected — decides every step |

The four continuous scenes above are still pending; `worm` never needed the fix.
If you consume observations/rewards only, every dataset here is usable as
published. If you need action labels — behavior cloning, action-conditioned
models, most offline RL — the discrete environments are now correct.

Because a collection run is not bit-reproducible (agents finish episodes
asynchronously, so the episode boundaries differ), the re-recorded tiers have
slightly different episode/transition counts and tier metrics than the files they
replaced — within ~1% and ~0.02 normalized respectively. The tables below carry
the new figures.

## Contents

| Dataset id | Episodes | Transitions | Observation | Action |
|---|---:|---:|---|---|
| `unity/crawler/expert-v0` | 1,048 | 1,000,002 | `Tuple(Box(126), Box(32))` | `Box(20, [-1, 1])` |
| `unity/crawler/medium-v0` | 1,129 | 1,000,000 | `Tuple(Box(126), Box(32))` | `Box(20, [-1, 1])` |
| `unity/crawler/simple-v0` | 1,344 | 1,000,000 | `Tuple(Box(126), Box(32))` | `Box(20, [-1, 1])` |
| `unity/pushblock/expert-v0` | 53,039 | 1,000,000 | `Tuple(Box(105), Box(105))` | `Discrete(7)` |
| `unity/pushblock/medium-v0` | 39,033 | 1,010,016 | `Tuple(Box(105), Box(105))` | `Discrete(7)` |
| `unity/pushblock/simple-v0` | 26,666 | 1,010,016 | `Tuple(Box(105), Box(105))` | `Discrete(7)` |
| `unity/soccer-twos/expert-v0` | 14,692 | 1,001,676 | ego `Dict` wrapping `Tuple(Box(264), Box(72))` | ego `Dict` wrapping `MultiDiscrete([3, 3, 3])` |
| `unity/soccer-twos/medium-v0` | 12,940 | 1,002,368 | ego `Dict` wrapping `Tuple(Box(264), Box(72))` | ego `Dict` wrapping `MultiDiscrete([3, 3, 3])` |
| `unity/soccer-twos/simple-v0` | 11,820 | 1,001,564 | ego `Dict` wrapping `Tuple(Box(264), Box(72))` | ego `Dict` wrapping `MultiDiscrete([3, 3, 3])` |
| `unity/dungeon-escape/expert-v0` | 35,505 | 1,007,206 | ego `Dict` wrapping `Tuple(Box(10), Box(360), Box(1))` | ego `Dict` wrapping `Discrete(7)` |
| `unity/dungeon-escape/medium-v0` | 28,314 | 1,012,457 | ego `Dict` wrapping `Tuple(Box(10), Box(360), Box(1))` | ego `Dict` wrapping `Discrete(7)` |
| `unity/dungeon-escape/simple-v0` | 22,542 | 1,001,594 | ego `Dict` wrapping `Tuple(Box(10), Box(360), Box(1))` | ego `Dict` wrapping `Discrete(7)` |
| `unity/3dball-hard/expert-v0` | 1,009 | 1,000,008 | `Tuple(Box(27), Box(18))` | `Box(2)` |
| `unity/3dball-hard/medium-v0` | 1,250 | 1,000,008 | `Tuple(Box(27), Box(18))` | `Box(2)` |
| `unity/3dball-hard/simple-v0` | 1,655 | 1,000,008 | `Tuple(Box(27), Box(18))` | `Box(2)` |
| `unity/pyramids/expert-v0` | 5,392 | 1,000,000 | `Tuple(Box(56), Box(56), Box(56), Box(4))` | `Discrete(5)` |
| `unity/pyramids/medium-v0` | 4,174 | 1,000,000 | `Tuple(Box(56), Box(56), Box(56), Box(4))` | `Discrete(5)` |
| `unity/pyramids/simple-v0` | 3,065 | 1,000,000 | `Tuple(Box(56), Box(56), Box(56), Box(4))` | `Discrete(5)` |
| `unity/worm/expert-v0` | 1,000 | 1,000,000 | `Box(64)` | `Box(9, [-1, 1])` |
| `unity/worm/medium-v0` | 1,000 | 1,000,000 | `Box(64)` | `Box(9, [-1, 1])` |
| `unity/worm/simple-v0` | 1,000 | 1,000,000 | `Box(64)` | `Box(9, [-1, 1])` |
| `unity/walker/expert-v0` | 1,458 | 1,000,010 | `Box(243)` | `Box(39, [-1, 1])` |
| `unity/walker/medium-v0` | 1,712 | 1,000,000 | `Box(243)` | `Box(39, [-1, 1])` |
| `unity/walker/simple-v0` | 2,247 | 1,000,000 | `Box(243)` | `Box(39, [-1, 1])` |

All environments, datasets, and stock policies use **ML-Agents `release_23`**.
Each dataset is stored at `<name>/<tier>/data/main_data.hdf5` with a sibling
`metadata.json` (`minari_version` 0.5.3). SoccerTwos and DungeonEscape use an
ego-agent schema — `observations["agents"]["agent_0"]`,
`actions["agents"]["agent_0"]` — one ego episode per physical agent; split by
`match_id` (see [docs/reproduction.md](docs/reproduction.md)).

## Loading

```python
from pathlib import Path
from huggingface_hub import snapshot_download
import minari

snapshot_download(
    repo_id="ccnets/causal-gpt-rl-unity-datasets",
    repo_type="dataset",
    allow_patterns="worm/**",                       # one env; drop to fetch all
    local_dir=Path.home() / ".minari" / "datasets" / "unity",
)
dataset = minari.load_dataset("unity/worm/expert-v0")
# calibrated tiers: minari.load_dataset("unity/worm/medium-v0" | ".../simple-v0")
print(dataset.observation_space, dataset.action_space)
```

## Quality ladders

**Why these tiers look the way they do.** In a normal RL pipeline, `medium`/`simple`
data would come from *early training checkpoints* on the way to the expert. We don't
have those checkpoints — only the final expert policy — so each tier is **reduced
backward from the expert**: instead of one constant perturbation (a single degraded
point), every episode draws its skill from a calibrated *range*, so the tier
reproduces the *distribution* of skill a checkpoint spread would have.

Each ladder is a monotone skill sequence built from ONE stock policy: `expert` is
the unmodified policy, `medium`/`simple` degrade it progressively. The degradation
is calibrated so the tier mean hits its normalized target while episodes cover a
continuous skill band (seed 2310000):

- **continuous** scenes (crawler, worm, walker, 3dball-hard) add Gaussian action
  noise, `noise_std` sampled per episode from a calibrated range;
- **discrete goal games** sample the policy's own action distribution via
  `softmax(logits/T)` on the exposed discrete logits — higher `T` picks plausible
  2nd/3rd-best actions (smooth, in-distribution degradation, unlike epsilon-style
  uniform-random swaps which are bimodal and off-distribution). PushBlock and
  Pyramids each sample `T` **per episode** from a calibrated range (each episode
  its own skill level, like a spread of early-training checkpoints);
- **cooperative** discrete DungeonEscape samples a softmax temperature `T` **per
  group per match** from a calibrated range — all three agents in a group share one
  `T`, so each match is one coherent skill level (like an early-training checkpoint);
  competitive self-play SoccerTwos samples a softmax temperature `T` **per team per
  match** from a (wider) calibrated range — the two teams draw independently, so each
  match pairs two skill levels (a random gap, like cross-checkpoint league play).

Normalization is `(candidate − random) / (expert − random)` on each env's tier
metric.

| env | tier metric | medium | simple | noise method |
|---|---|---:|---:|---|
| `crawler` | episode return | 0.82 | 0.60 | per-episode `noise_std` range |
| `worm` | episode return | 0.80 | 0.60 | per-episode `noise_std` range |
| `walker` | episode return | 0.82 | 0.61 | per-episode `noise_std` range |
| `3dball-hard` | episode return | 0.80 | 0.60 | per-episode `noise_std` range |
| `pushblock` | step-reward | 0.80 | 0.60 | per-episode softmax temperature range |
| `pyramids` | step-reward | 0.79 | 0.60 | per-episode softmax temperature range |
| `dungeon-escape` | step-reward | 0.80 | 0.61 | per-group softmax temperature range |
| `soccer-twos` | match score | ≈0.80 | ≈0.60 | per-team softmax temperature range |

`soccer-twos` figures are approximate: self-play stays balanced, so tier skill
cannot be read off the shipped returns and comes from side-swapped calibration.
Sparse-reward scenes (pyramids, pushblock, dungeon, soccer) have a physically
bimodal per-episode outcome (solve vs. time-out), so the episode-return histogram
barely separates the tiers. For the goal games (`pushblock`, `pyramids`, `dungeon-escape`) the tier metric is
therefore **mean step-reward** = `return / episode length`, which grades *how
efficiently* the goal is reached — a continuous skill signal — normalized
`expert = 1 / random = 0` against the shipped `expert-v0` step-reward anchor
(pushblock 0.3537, pyramids 0.01263, dungeon-escape 0.01936). `simple` targets 0.60
and the shipped tiers all sit at **0.60–0.61** (SoccerTwos, scored by side-swapped match
score rather than step-reward, is approximate); `medium` spans 0.79–0.82.

"termination rate" below = fraction of episodes ending on the env's terminal
condition — a fall for Crawler/Walker/3DBallHard, a solved push for PushBlock,
maze solve ("reach") for Pyramids; Worm never terminates (truncated at 1000).

<details>
<summary><b>Per-environment ladder tables</b> — exact ranges, anchors, mean returns, rates</summary>

**Crawler** — anchors: expert 2576.4899, random −0.88

| tier | norm. return | noise_std range | mean return | term. rate | mean ep length |
|---|---:|---:|---:|---:|---:|
| `expert-v0` | 1.00 | -- | 2576.4899 | 0.077 | 954.20 |
| `medium-v0` | ≈0.82 | 0.15–0.23 | 2113.0960 | 0.209 | 885.74 |
| `simple-v0` | ≈0.60 | 0.21–0.32 | 1557.9818 | 0.460 | 744.05 |

**Worm** — anchors: expert 1044.1063, random 0.80 (never terminates)

| tier | norm. return | noise_std range | mean return | term. rate | mean ep length |
|---|---:|---:|---:|---:|---:|
| `expert-v0` | 1.00 | -- | 1044.1063 | 0.000 | 1000.00 |
| `medium-v0` | ≈0.80 | 0.09–0.15 | 839.4910 | 0.000 | 1000.00 |
| `simple-v0` | ≈0.60 | 0.15–0.22 | 632.4776 | 0.000 | 1000.00 |

**Walker** — anchors: expert 1363.6454, random −0.49 (high intrinsic return variance)

| tier | norm. return | noise_std range | mean return | term. rate | mean ep length |
|---|---:|---:|---:|---:|---:|
| `expert-v0` | 1.00 | -- | 1363.6454 | 0.512 | 685.88 |
| `medium-v0` | ≈0.82 | 0.05–0.11 | 1122.0456 | 0.652 | 584.11 |
| `simple-v0` | ≈0.61 | 0.09–0.15 | 835.2383 | 0.822 | 445.04 |

**3DBallHard** — anchors: expert 99.1077, random 0.84 (capped balance: near-bimodal)

| tier | norm. return | noise_std range | mean return | term. rate | mean ep length |
|---|---:|---:|---:|---:|---:|
| `expert-v0` | 1.00 | -- | 99.1077 | 0.001 | 991.09 |
| `medium-v0` | ≈0.80 | 0.28–0.38 | 79.6258 | 0.341 | 800.01 |
| `simple-v0` | ≈0.60 | 0.31–0.44 | 59.7502 | 0.612 | 604.23 |

**PushBlock** — tier metric is **mean step-reward** (`return / length`); step-reward anchors: expert 0.353699 (shipped `expert-v0`, 53,039 ep), random ≈−0.0005 (soft). Each tier samples `T` **per episode** from a calibrated range (each episode its own skill level, like a spread of early-training checkpoints). Higher `T` still solves but wanders more, so the return stays high (≈4.9) while step-reward falls with efficiency. term. rate = solved.

| tier | norm. step-reward | `T` (per episode) | mean step-reward | mean return | solved rate | mean ep length |
|---|---:|---:|---:|---:|---:|---:|
| `expert-v0` | 1.00 | -- | 0.353699 | 4.972780 | ≈1.00 | 18.85 |
| `medium-v0` | 0.80 | `U(2.1, 3.2)` | 0.282037 | 4.961591 | 0.997 | 25.88 |
| `simple-v0` | 0.60 | `U(3.25, 4.75)` | 0.213119 | 4.942463 | 0.996 | 37.88 |

**Pyramids** — tier metric is **mean step-reward** (`return / length`); step-reward anchors: expert 0.012633 (shipped `expert-v0`, 5,392 ep), random −0.001 (sparse: policy always times out at random). Each tier samples `T` **per episode** from a calibrated range (each episode its own skill level, like a spread of early-training checkpoints). Higher `T` reaches the goal less directly (longer episodes) so step-reward falls while return stays near expert. term. rate = mazes reached/solved.

| tier | norm. step-reward | `T` (per episode) | mean step-reward | mean return | reach rate | mean ep length |
|---|---:|---:|---:|---:|---:|---:|
| `expert-v0` | 1.00 | -- | 0.012633 | 1.797730 | 0.9911 | 185.46 |
| `medium-v0` | 0.79 | `U(2.7, 4.9)` | 0.009784 | 1.742249 | 0.990 | 239.58 |
| `simple-v0` | 0.60 | `U(4.8, 7.1)` | 0.007201 | 1.636223 | 0.981 | 326.26 |

**DungeonEscape** — cooperative goal game; tier metric is **mean step-reward** (`return / length`); step-reward anchors: expert 0.019355 (shipped `expert-v0`, 35,505 ego ep), random ≈0.00015. A softmax temperature `T` is sampled **per cooperative group per match** from a calibrated range and shared by all three agents (one coherent skill level per match, like an early-training checkpoint); higher `T` reaches the exit less efficiently so step-reward falls while the group still often succeeds. Group success `any(agent_return > 0)` (`group_metadata.jsonl`) therefore stays high across tiers and no longer separates them — step-reward does.

| tier | norm. step-reward | `T` range (per group) | mean step-reward | group success | mean ep length |
|---|---:|---|---:|---:|---:|
| `expert-v0` | 1.00 | stock (no noise) | 0.019355 | 0.9604 | 28.37 |
| `medium-v0` | 0.80 | `U(1.5, 3.0)` | 0.015446 | 0.9183 | 35.76 |
| `simple-v0` | 0.61 | `U(3.0, 4.0)` | 0.011950 | 0.8585 | 44.43 |

**SoccerTwos** — tier metric is **side-swapped match score** vs the fixed stock policy (self-play, so mean shipped return ≈ −0.04 for both tiers and does not track skill; figures are calibration estimates). Each **team** draws a softmax temperature `T` **independently per match** from a (wide) calibrated range on the exposed MultiDiscrete logits, shared by its two agents — so each match pairs two independently-drawn skill levels (a random gap = cross-checkpoint league play). A tier is set by the ego team's own drawn `T`; the norm is that band's mean side-swapped match score vs the fixed expert. term. rate = matches always terminate.

| tier | approx. norm. skill | `T` (per team/match) | matches | mean ep length |
|---|---:|---|---:|---:|
| `expert-v0` | 1.00 | stock (no noise) | 3,673 | 68.18 |
| `medium-v0` | ≈0.80 | `U(1.3, 2.6)` | 3,235 | 77.46 |
| `simple-v0` | ≈0.60 | `U(2.1, 3.4)` | 2,955 | 84.73 |

</details>

**Build & reproduction steps** (per-env recording recipe): see
[docs/reproduction.md](docs/reproduction.md).

## Provenance & attribution

- Trajectories were generated by running Unity ML-Agents material (the builds and
  stock policies are Apache-2.0; see the envs repo for provenance and licenses).
- The **trajectory data in this repo** is an original recording licensed
  **CC-BY-4.0**. Please attribute *ccnets — Causal GPT-RL* and note the Unity
  ML-Agents source environment.

## License

Creative Commons Attribution 4.0 International (CC-BY-4.0). See `LICENSE`.