ccnets/causal-gpt-rl-unity-envs

Causal GPT-RL — Unity ML-Agents environments

Public materials for running Causal GPT-RL policies in Unity ML-Agents. Each environment is a model-removed Unity build — the engine binary with no baked-in policy; the policy is supplied at run time. All eight environments are paired with their stock built-in policies, and the four discrete-action ones additionally ship a logits-exposing variant under tier_policies/.

Builds are published for Windows x86-64 and, under linux/, for Linux x86-64. No macOS builds are published.

Environments

Environment	Build	Stock policy	Agents	Observation	Action
Crawler	Crawler/	Crawler.onnx	10	158 (126 + 32)	Box(20) continuous
3DBallHard	3DBallHard/	3DBallHard.onnx	12	45 (27 + 18)	Box(2) continuous
PushBlock	PushBlock/	PushBlock.onnx	32	210 (105 + 105)	Discrete(7)
DungeonEscape	DungeonEscape/	DungeonEscape.onnx	36	371	Discrete(7)
SoccerTwos	SoccerTwos/	SoccerTwos.onnx	32	336 (264 + 72)	MultiDiscrete([3,3,3])
Pyramids	Pyramids/	Pyramids.onnx	16	172 (56 + 56 + 56 + 4)	Discrete(5)
Worm	Worm/	Worm.onnx	10	64	Box(9) continuous
Walker	Walker/	Walker.onnx	10	243	Box(39) continuous

Companion repos

• Policy (ONNX): ccnets/causal-gpt-rl-unity
• Recorded trajectories + reproduction recipe: ccnets/causal-gpt-rl-unity-datasets

Reproduce — record trajectories from these builds with the collection & measurement recipe: examples/unity_collection.

Getting a build

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="ccnets/causal-gpt-rl-unity-envs",
    repo_type="dataset",
    allow_patterns=["Crawler/**", "Crawler.onnx"],   # one environment, Windows
    local_dir="unity-envs",
)

One environment is about 95 MB. The whole repo is 1.6 GB — eight environments across both platforms — so fetch the ones you need. Linux builds are under linux/, as linux/Crawler/**; the logits-exposing policy variants are under tier_policies/.

Layout

Every environment is one folder with its stock policy beside it:

Crawler/                 # model-removed build: .exe, _Data/, UnityPlayer.dll
Crawler.onnx             # stock built-in policy (ml-agents-release_23)

DungeonEscape and SoccerTwos name their executable UnityEnvironment.exe rather than <Environment>.exe. linux/ mirrors the same tree for Linux x86-64. tier_policies/ holds logits-exposing variants of the four discrete policies, with the calibrated temperature bands in its own README.

Each build is model-removed: no policy is baked in — supply one at run time (the stock policy here, or the Causal GPT-RL policy from the model repo). All launch headless.

Crawler

• Scene: 10 agents, observation 158-d (sensor channels 126 + 32), action Box(20) continuous, behavior Crawler?team=0. In the companion dataset the observation is stored as Tuple(Box(126), Box(32)).
• Built from the customized ML-Agents release-23 project with the baked model removed. Crawler.onnx is the unmodified release-23 stock policy used as the collection driver; size 2,472,183 bytes, SHA-256 df5dd19f28e89678f96b794e911564c21d0039160e88d99fe04f3939442ffdd7.

3DBallHard

• Scene: 12 agents, observation 45-d (sensor channels 27 + 18), action Box(2) continuous, behavior 3DBallHard?team=0.
• Built from the customized ML-Agents release-23 project with the baked model removed. 3DBallHard.onnx is the unmodified release-23 stock policy used as the collection driver; SHA-256 380aed53f8bcfd3b42a7eb758923c2eb7a98abe91c44b44ba2288924e65de701.
• The ONNX takes the two sensor channels as separate inputs — obs_0 [27] and obs_1 [18], not one concatenated [45] vector. It returns continuous_actions (sampled) and deterministic_continuous_actions (greedy), both float32[B, 2].

PushBlock

• Scene: 32 agents, observation 210-d (two sensors 105 + 105), action a single Discrete(7) branch (0 = no-op; 1–6 = move/turn), behavior PushBlock?team=0. In the companion dataset the observation is stored as Tuple(Box(105), Box(105)).
• Built from the customized ML-Agents release-23 project with the baked model removed. PushBlock.onnx is the unmodified release-23 stock policy used as the collection driver; SHA-256 1c8230f87d3067343f8b8b8ebf0839843c855d4d7912111546a7522e30e4bfbe.
• The stock ONNX emits Discrete(7) actions for external Python control of the model-removed build.

DungeonEscape

• Scene: 36 agents arranged as 12 arenas x 3 cooperative agents, observation 371-d, action a single Discrete(7) branch.
• Built from the customized ML-Agents release-23 project with the baked model removed. DungeonEscape.onnx is the release-23 built-in policy used as the collection driver; SHA-256 141e3595dbaa3ea369b2c5599e6e750b53b8a9c41414a6b5c7043617860691ab.
• The three agents in each arena share a cooperative objective while each policy invocation receives its own ego observation.
• Evaluation tutorial: examples/unity.

SoccerTwos

• Scene: 32 agents arranged as eight 2-vs-2 fields, observation channels 264 + 72 = 336, action MultiDiscrete([3,3,3]); behaviors are SoccerTwos?team=0 and SoccerTwos?team=1 with 16 agents each.
• Built from the customized ML-Agents release-23 project with the baked model removed. SoccerTwos.onnx is the release-23 stock policy used as the public opponent and collection driver.
• The stock graph is computation-equivalent to the release-23 source policy; only dangling output declarations that reference no graph values were removed so current ONNX Runtime versions can load it.
• Causal-vs-stock tutorial: examples/unity.

Pyramids

• Scene: 16 agents, observation channels 56 + 56 + 56 + 4 = 172, action Discrete(5). The stock release-23 policy is redistributed unmodified; SHA-256 0ab9ec55f788018cd0b01ebe28b198b4714e85f27f48710d1ebc27fd455a0e07.

Worm

• Scene: 10 agents, one 64-d observation sensor and continuous Box(9) action. The stock release-23 policy is redistributed unmodified; SHA-256 a5dd14742d9e3d1d1ca4840bca600b669b37d630977091ee10286ed7a833411f.

Walker

• Scene: 10 agents, one 243-d observation sensor and continuous Box(39) action. The stock release-23 policy is redistributed unmodified; SHA-256 28f675937108a850e5c3ea2894a78552f51290c0d6220ebf3ee2c223d016c434.

Running an unsigned build (Windows)

These builds are not code-signed, so Windows SmartScreen may warn on first launch ("Windows protected your PC" → More info → Run anyway). Verify the download before running.

Provenance & scope

• The builds were produced from customized ML-Agents projects. All eight — Crawler, 3DBallHard, PushBlock, Pyramids, Worm, Walker, DungeonEscape, and SoccerTwos — derive from ml-agents-release_23. The distributed stock policies match their respective build contracts. Other example scenes are not included.
• Not source-rebuildable from stock ML-Agents: the project changes live in the compiled binaries. The builds are provided as-is.
• Excluded from each build: its *_BurstDebugInformation_DoNotShip/ folder (Burst debug symbols, not for shipping).

License

The Unity builds and the stock policies (Crawler.onnx, PushBlock.onnx, Pyramids.onnx, Worm.onnx, Walker.onnx, 3DBallHard.onnx, SoccerTwos.onnx, DungeonEscape.onnx) are Apache-2.0 (Unity ML-Agents), as are the tier_policies/ exports derived from them. See LICENSE and NOTICE. The builds are modified derivatives — the modifications are described in NOTICE. The Unity runtime components (UnityPlayer.dll, UnityCrashHandler64.exe, MonoBleedingEdge/) are redistributed as part of a built player, subject to Unity Technologies' software terms.