Upload README.md with huggingface_hub

README.md

@@ -21,17 +21,17 @@ size_categories:
 
 > **Dataset** (you are here) · **[Showcase](https://huggingface.co/spaces/exptech/g1-moves)** (interactive gallery) · **[Code](https://github.com/experientialtech/g1-moves)** (scripts & docs)
 
-60 motion capture clips for the Unitree G1 humanoid robot (29 DOF, mode 15), captured with [MOVIN TRACIN](https://movin3d.com/) markerless motion capture and [video2robot](https://github.com/experientialtech/video2robot) monocular video extraction. Each clip is provided at multiple pipeline stages: raw mocap (BVH/FBX), retargeted robot joint trajectories (PKL), processed RL training data (NPZ), and trained ONNX policies with baked-in observation normalization.
+60 motion capture clips for the Unitree G1 humanoid robot (edition EDU, 29 DOF), captured from real performers in Austin, TX using [MOVIN TRACIN](https://movin3d.com/) markerless motion capture and [video2robot](https://github.com/experientialtech/video2robot) monocular video extraction. Each clip is provided at multiple pipeline stages: raw mocap (BVH/FBX), retargeted robot joint trajectories (PKL), processed RL training data (NPZ), and trained ONNX policies with baked-in observation normalization. All 60 clips have trained RL policies and ONNX exports ready for deployment.
 
 ## Dataset Summary
 
 | | |
 |---|---|
 | **Total clips** | 60 |
-| **Total duration** | 29.5 minutes (106,149 frames at 60 FPS) |
+| **Total duration** | 29.6 minutes (106,434 frames at 60 FPS) |
 | **Categories** | Dance (28), Karate (27), Bonus (5) |
 | **Clip duration** | 6.5s - 119.5s |
-| **Robot** | Unitree G1, mode 15, 29 DOF |
+| **Robot** | Unitree G1, edition EDU, 29 DOF |
 | **Capture system** | MOVIN TRACIN (markerless, LiDAR + vision) |
 | **Retarget SDK** | [movin_sdk_python](https://github.com/MOVIN3D/movin_sdk_python) |
 
@@ -65,8 +65,9 @@ Each clip lives in its own subfolder organized by pipeline stage:
     <clip>.npz            Training data with forward kinematics
     <clip>_training.gif   Training visualization
     <clip>_training.mp4   Training visualization video
-  policy/                 Trained RL policy (if available)
+  policy/                 Trained RL policy
     <clip>_policy.pt      PyTorch checkpoint
+    <clip>_policy.onnx    ONNX export (baked-in observation normalization)
     <clip>_policy.gif     Policy rollout animation
     <clip>_policy.mp4     Policy rollout video
     agent.yaml            PPO hyperparameters
@@ -223,6 +224,7 @@ print("Least energetic:", by_energy[-1][0])
 | `B_` | [Mitch Chaiet](https://mitchchaiet.com/) | Bonus clips + some dance |
 | `J_` | [Jasmine Coro](https://jasminecoro.com/) | Dance choreography |
 | `M_` | [Mike Gassaway](https://www.backstage.com/u/mike-gassaway/) | Karate / martial arts |
+| `V_` | [Maya Coro](https://www.backstage.com/u/maya-coro/) | Fencing (via video2robot) |
 
 ### Processing Pipeline
 
@@ -230,7 +232,7 @@ print("Least energetic:", by_energy[-1][0])
 2. **Retarget**: [movin_sdk_python](https://github.com/MOVIN3D/movin_sdk_python) maps human skeleton to G1 joint limits (1.75m human height)
 3. **Ground calibration**: MuJoCo forward kinematics finds minimum foot Z, shifts root for ground contact
 4. **Training data**: MuJoCo computes full-body forward kinematics (positions, orientations, velocities)
-5. **RL training**: PPO with motion imitation rewards in MuJoCo-Warp (4096 parallel envs)
+5. **RL training**: PPO with motion imitation rewards in MuJoCo-Warp (8,192 parallel envs)
 
 ## Metadata Files
 
@@ -242,16 +244,16 @@ print("Least energetic:", by_energy[-1][0])
 
 ## Training Quality
 
-44 of 60 clips have trained RL policies. Aggregate metrics across all trained policies:
+All 60 clips have trained RL policies and ONNX exports. Aggregate metrics across all trained policies:
 
 | Metric | Mean | Min | Max |
 |--------|------|-----|-----|
-| Reward | 39.7 | 33.1 | 45.6 |
-| Episode length | 493 / 500 | 458 | 500 |
-| Body position error | 0.045 m | 0.031 m | 0.078 m |
-| Joint position error | 0.56 rad | 0.42 rad | 0.85 rad |
+| Reward | 40.1 | 20.6 | 45.6 |
+| Episode length | 493 / 500 | 419 | 500 |
+| Body position error | 0.044 m | 0.030 m | 0.108 m |
+| Joint position error | 0.54 rad | 0.39 rad | 0.93 rad |
 
-Each policy directory contains a `training_log.csv` with per-iteration metrics (rewards, losses, tracking errors, termination rates).
+Each policy directory contains a `training_log.csv` with per-iteration metrics (rewards, losses, tracking errors, termination rates). ONNX exports include baked-in observation normalization (160 obs → [512, 256, 128] → 29 actions) for direct deployment without the training framework.
 
 ## Standalone Inference