Add discrete PushBlock expert Minari dataset

Converted from the ML-Agents release_20 ExpertPushBlock.demo. Adds 96 episodes / 2,557 transitions with Tuple(Box(105), Box(105)) observations and Discrete(7) actions.

.gitattributes

@@ -63,3 +63,4 @@ crawler/simple-v0/data/main_data.hdf5 filter=lfs diff=lfs merge=lfs -text
 crawler/medium-v0/data/main_data.hdf5 filter=lfs diff=lfs merge=lfs -text
 pushblock-with-input/expert-v0/data/main_data.hdf5 filter=lfs diff=lfs merge=lfs -text
 3dball-hard/expert-v0/data/main_data.hdf5 filter=lfs diff=lfs merge=lfs -text
+pushblock/expert-v0/data/main_data.hdf5 filter=lfs diff=lfs merge=lfs -text

README.md

@@ -9,14 +9,9 @@ tags:
 
 # Causal GPT-RL — Unity ML-Agents trajectories
 
-Recorded **Crawler trajectories**, collected by running the stock built-in policy
-in a public Unity build and packaged as a [Minari](https://minari.farama.org)
-dataset.
-
-- **Dataset id:** `unity/crawler/expert-v0`
-- **Observation:** flat `Box(158)` · **Action:** `Box(20, [-1, 1])` continuous —
-  the same env-less flat convention as the Gymnasium / MuJoCo Minari datasets, so
-  it consumes identically.
+Recorded Unity ML-Agents trajectories packaged as
+[Minari](https://minari.farama.org) datasets. The collection includes continuous,
+discrete, and hybrid action spaces for testing Causal GPT-RL across space types.
 
 **Companion repos**
 - Unity build + stock policy:
@@ -25,10 +20,31 @@ dataset.
 
 ## Contents
 
-- **1,047 episodes · 1,000,000 transitions.** Minari HDF5 at
-  `crawler/expert-v0/data/main_data.hdf5` (+ `metadata.json`), plus
-  `crawler/namespace_metadata.json` for the `unity` namespace.
-- Observation `Box(158)`, action `Box(20, [-1, 1])`, `minari_version` 0.5.3.
+| Dataset id | Episodes | Transitions | Observation | Action |
+|---|---:|---:|---|---|
+| `unity/crawler/expert-v0` | 1,047 | 1,000,000 | `Box(158)` | `Box(20, [-1, 1])` |
+| `unity/crawler/medium-v0` | — | 1,000,000 | `Box(158)` | `Box(20, [-1, 1])` |
+| `unity/crawler/simple-v0` | — | 1,000,000 | `Box(158)` | `Box(20, [-1, 1])` |
+| `unity/pushblock-with-input/expert-v0` | 1,089 | 1,000,000 | `Tuple(Box(105), Box(105))` | `Tuple(Box(2), Discrete(2))` |
+| `unity/pushblock/expert-v0` | 96 | 2,557 | `Tuple(Box(105), Box(105))` | `Discrete(7)` |
+| `unity/3dball-hard/expert-v0` | — | — | See metadata | See metadata |
+
+Every dataset is stored at `<name>/<tier>/data/main_data.hdf5` with a sibling
+`metadata.json` and uses `minari_version` 0.5.3. PushBlock is converted from the
+official release_20 `ExpertPushBlock.demo`: 95 episodes terminate normally and
+the final recording fragment is represented as one truncated episode.
+
+Load the discrete PushBlock dataset on another machine with:
+
+```python
+import minari
+
+dataset = minari.load_dataset(
+    "unity/pushblock/expert-v0", download=True
+)
+print(dataset.observation_space)  # Tuple(Box(105), Box(105))
+print(dataset.action_space)       # Discrete(7)
+```
 
 The recipe below regenerates this dataset from public materials at any time.
 
@@ -47,8 +63,9 @@ Everything needed to regenerate this dataset from public materials only:
    agent's per-step `(observation, action, reward, termination)` to per-episode
    files. The Crawler scene runs 10 agents on a shared decision period, with
    action-repeat gaps and independent per-agent terminations.
-3. **Package to Minari.** Flatten the sensor channels into `Box(158)`, keep
-   actions as `Box(20)`, and write the Minari dataset `unity/crawler/expert-v0`.
+3. **Package to Minari.** Preserve the spaces declared by each build. For the
+   legacy PushBlock demonstration, parse the ML-Agents protobuf records using the
+   release_20 transition alignment and package its single branch as `Discrete(7)`.
 
 The recipe ends at the Minari dataset — a portable, env-less trajectory set.
 

pushblock/expert-v0/data/main_data.hdf5

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aef2791a97b925757c2411d2cda7734971b7f5095e2524e53fa63b13ca0a34df
+size 3970632

pushblock/expert-v0/data/metadata.json

@@ -0,0 +1 @@
+{"total_episodes": 96, "total_steps": 2557, "data_format": "hdf5", "observation_space": "{\"type\": \"Tuple\", \"subspaces\": [{\"type\": \"Box\", \"dtype\": \"float32\", \"shape\": [105], \"low\": [-Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity], \"high\": [Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity]}, {\"type\": \"Box\", \"dtype\": \"float32\", \"shape\": [105], \"low\": [-Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity, -Infinity], \"high\": [Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity, Infinity]}]}", "action_space": "{\"type\": \"Discrete\", \"dtype\": \"int64\", \"start\": 0, \"n\": 7}", "dataset_id": "unity/pushblock/expert-v0", "author": ["junho"], "description": "ML-Agents PushBlock expert demonstration from release_20 (legacy Barracuda policy, Discrete(7)).", "minari_version": "0.5.3", "dataset_size": 4.0}

pushblock/namespace_metadata.json

@@ -0,0 +1 @@
+{}