Yang commited on
Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,61 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
tags:
|
| 4 |
+
- robotics
|
| 5 |
+
- inverse-dynamics-model
|
| 6 |
+
- idm
|
| 7 |
+
- omx
|
| 8 |
+
- gr00t
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# IDM-OMX-v3-50K
|
| 12 |
+
|
| 13 |
+
NVIDIA GR00T IDM (Inverse Dynamics Model) fine-tuned for ROBOTIS OMX-F robot.
|
| 14 |
+
|
| 15 |
+
## Model Details
|
| 16 |
+
|
| 17 |
+
- **Architecture**: SiGLIP2(768d) → VL Transformer(4L) → DiT(8L, 1024d)
|
| 18 |
+
- **Parameters**: 612M
|
| 19 |
+
- **Embodiment**: OMX (slot 30)
|
| 20 |
+
- **Action Space**: 6-DOF delta joint angles (radians), 30 FPS
|
| 21 |
+
- **Action Horizon**: 16 steps
|
| 22 |
+
|
| 23 |
+
## Training
|
| 24 |
+
|
| 25 |
+
- **Base Model**: IDM-OMX-v2 (20K steps checkpoint)
|
| 26 |
+
- **Dataset**: 558 episodes, 257,466 frames (OMX combined + SO-101 megamix FK→IK converted)
|
| 27 |
+
- **Steps**: 50,000 (6.21 epochs)
|
| 28 |
+
- **Batch Size**: 16 (effective 32 with gradient accumulation)
|
| 29 |
+
- **Learning Rate**: 1e-4 (cosine schedule)
|
| 30 |
+
- **Duration**: 5h 08min on RTX 5090
|
| 31 |
+
- **Final Train Loss**: 0.2644
|
| 32 |
+
|
| 33 |
+
## Evaluation
|
| 34 |
+
|
| 35 |
+
| Metric | Baseline (10K) | v2 (20K) | **v3 (50K)** |
|
| 36 |
+
|--------|:-:|:-:|:-:|
|
| 37 |
+
| Normalized MAE | 0.6092 | 0.5582 | **0.3551 (-36.4%)** |
|
| 38 |
+
| Forward Loss | 0.3942 | 0.3453 | **0.2517 (-27.1%)** |
|
| 39 |
+
|
| 40 |
+
### Per-Joint MAE (degrees)
|
| 41 |
+
|
| 42 |
+
| Joint | v2 | v3 | Change |
|
| 43 |
+
|-------|---:|---:|---:|
|
| 44 |
+
| shoulder_pan | 82.12 | 71.76 | -12.6% |
|
| 45 |
+
| shoulder_lift | 83.01 | 72.74 | -12.4% |
|
| 46 |
+
| elbow_flex | 65.65 | 30.12 | **-54.1%** |
|
| 47 |
+
| wrist_flex | 141.40 | 96.43 | **-31.8%** |
|
| 48 |
+
| wrist_roll | 67.87 | 56.43 | -16.9% |
|
| 49 |
+
| gripper | 6.50 | 4.21 | -35.2% |
|
| 50 |
+
|
| 51 |
+
## Usage
|
| 52 |
+
|
| 53 |
+
|
| 54 |
+
|
| 55 |
+
### Pseudo-label Extraction
|
| 56 |
+
|
| 57 |
+
|
| 58 |
+
|
| 59 |
+
## Citation
|
| 60 |
+
|
| 61 |
+
Based on [NVIDIA GR00T IDM](https://github.com/NVIDIA-Omniverse/GR00T-Dreams).
|