docs/additional-simulations.md

Run the additional model-removed simulations

Released 2026-10-22: PushBlockCollab, FoodCollector, StrikersVsGoalie, Hallway, GridWorld, Sorter, WallJump. The players contain no baked-in policy. Download their stock ONNX separately or send your own actions through ML-Agents.

Use a dedicated Python 3.10 environment:

pip install -r unity-envs/tools/requirements-simulation.txt

Download one Windows environment, external policy, and runner:

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="ccnets/causal-gpt-rl-unity-envs", repo_type="dataset",
    allow_patterns=["PushBlockCollab/**", "PushBlockCollab.onnx", "tools/**"],
    local_dir="unity-envs",
)

python unity-envs/tools/probe_simulation.py --build unity-envs/PushBlockCollab/PushBlockCollab.exe --policies unity-envs --report probe.json --ticks 1000

Add --graphics --time-scale 1 to view the scene in real time. Use --random to supply random actions without loading any ONNX. Set distinct --worker-id values when launching multiple players concurrently. Close the run with Ctrl+C.

For FoodCollector, fetch FoodCollector/** and GridFoodCollector.onnx instead. For StrikersVsGoalie, fetch StrikersVsGoalie/**, Striker.onnx, and Goalie.onnx; both role policies are required. The runner reads policy names from each player's simulation-build.json and routes them to the corresponding behavior.

Linux players are under linux/<Environment>/. Download that prefix instead of the Windows prefix, use the same root ONNX files, run chmod +x unity-envs/linux/<Environment>/<Environment>.x86_64, and pass the .x86_64 path to --build. A graphical run needs an available display. These Linux players were startup-tested under Ubuntu/WSL; stock-policy rollouts were verified on Windows only.

The runner preserves the native grid tensor and separate vector sensors, applies action masks, and handles continuous, discrete, and hybrid action spaces. It is a smoke-test/viewer tool and does not record Minari trajectories or instruction labels. The JSON reports contain raw reward sums, not calibrated skill scores; group-reward sums count rewards on each agent row, not unique team outcomes.

tools/AdditionalSimulationBuild.cs is the build helper. It opens copies of the three release-23 scenes in an isolated Unity project, unpacks prefabs, removes model references, sets external-control behavior, and rejects a build if its packed asset report includes ONNX/NN. Run with Unity 6000.0.40f1:

-batchmode -force-d3d11 -projectPath <isolated-release23-project> -executeMethod AdditionalSimulationBuild.Build -simulationOutput <output>

Place the helper in Assets/Editor/ first. For Linux add -buildTarget Linux64 -simulationTarget StandaloneLinux64. The local customized source project used here is not redistributed; this helper alone is not a full source project archive.

Remaining four scenes

Download <Environment>/**, its policy file(s), and tools/** as above:

Environment	External files	Runtime requirement
Hallway	Hallway.onnx	Per-agent recurrent memory, reset on episode endings
GridWorld	GridWorld.onnx	Camera rendering and action masks
Sorter	Sorter.onnx	Preserve padded variable-size buffer
WallJump	SmallWallJump.onnx, BigWallJump.onnx	Route dynamic behavior names to both policies

python unity-envs/tools/probe_simulation.py --build unity-envs/Hallway/Hallway.exe --policies unity-envs --report hallway-probe.json --ticks 1000
python unity-envs/tools/probe_simulation.py --build unity-envs/GridWorld/GridWorld.exe --policies unity-envs --report grid-probe.json --ticks 1000 --render
python unity-envs/tools/probe_simulation.py --build unity-envs/Sorter/Sorter.exe --policies unity-envs --report sorter-probe.json --ticks 1000
python unity-envs/tools/probe_simulation.py --build unity-envs/WallJump/WallJump.exe --policies unity-envs --report wall-probe.json --ticks 1000

Hallway's state is keyed by agent ID, not batch row. Memory is advanced only when an agent requests a decision and cleared on terminal or interrupted steps. The runner registers StatsSideChannel for scene statistics. GridWorld uses a rendering device even in --render batchmode; a Linux server must provide a working graphics backend/display for camera observations. Linux startup checks alone do not validate camera pixels. Sorter exposes VARIABLE_SIZE on its buffer dimension; do not discard this contract when designing a recorder.

Wall Jump's existing controller keeps its per-episode behavior switch, with all model references removed. Both behavior specs were exercised in the Windows probe; their initial allocation was 8 SmallWallJump and 16 BigWallJump agents, and this allocation changes during play. Count physical agents, not cumulative behavior slots, for scene population. All four Windows stock probes passed 1,000 ticks. Linux policy rollouts remain untested; native startup checks passed.

The build helper now supports all seven scenes. Add -simulationEnvironments Hallway,GridWorld,Sorter,WallJump to build only this batch. Some scenes need a graphics device while building their lighting data; keep the editor in batchmode and omit -nographics for those builds.