Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,52 @@
|
|
| 1 |
-
---
|
| 2 |
-
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: opentau
|
| 3 |
+
tags:
|
| 4 |
+
- robotics
|
| 5 |
+
- vla
|
| 6 |
+
- pi05
|
| 7 |
+
- libero
|
| 8 |
+
- manipulation
|
| 9 |
+
- flow-matching
|
| 10 |
+
- pytorch
|
| 11 |
+
base_model: williamyue/pi05_base
|
| 12 |
+
license: apache-2.0
|
| 13 |
+
datasets:
|
| 14 |
+
- physical-intelligence/libero
|
| 15 |
+
repo_url: https://github.com/TensorAuto/OpenTau
|
| 16 |
+
---
|
| 17 |
+
|
| 18 |
+
# pi0.5-libero-finetune
|
| 19 |
+
|
| 20 |
+
A **pi0.5 (π₀.₅)** Vision-Language-Action (VLA) model, finetuned on the **LIBERO** robotic manipulation benchmark. This model is designed to follow natural language instructions to perform manipulation tasks in a tabletop environment, leveraging the OpenTau framework.
|
| 21 |
+
|
| 22 |
+
**For full documentation, evaluation results, and inference code, please visit the repository:**
|
| 23 |
+
<br>
|
| 24 |
+
👉 **[https://github.com/TensorAuto/OpenTau](https://github.com/TensorAuto/OpenTau)**
|
| 25 |
+
|
| 26 |
+
---
|
| 27 |
+
|
| 28 |
+
## Model Details
|
| 29 |
+
|
| 30 |
+
### Description
|
| 31 |
+
- **Model Type:** Vision-Language-Action (VLA) Model
|
| 32 |
+
- **Base Architecture:** π₀.₅ (pi0.5) by Physical Intelligence
|
| 33 |
+
- **Backbone:** PaliGemma-2B (VLM) + Gemma-300M (Action Expert)
|
| 34 |
+
- **Training Data:** LIBERO (Lifelong Robot Learning) Benchmark
|
| 35 |
+
- **Framework:** OpenTau
|
| 36 |
+
|
| 37 |
+
### Architecture
|
| 38 |
+
The pi0.5 architecture uses a flow-matching-based policy designed for open-world generalization. It combines a Visual Language Model (VLM) for high-level semantic understanding with a smaller "action expert" model that generates continuous joint trajectories (50-step action chunks) via flow matching.
|
| 39 |
+
|
| 40 |
+
---
|
| 41 |
+
|
| 42 |
+
## Training and Evaluation
|
| 43 |
+
|
| 44 |
+
### Dataset
|
| 45 |
+
This model was finetuned on the **LIBERO** benchmark dataset. The LIBERO suite consists of human-teleoperated demonstrations for tabletop manipulation, covering:
|
| 46 |
+
- **Spatial Generalization** (libero_spatial)
|
| 47 |
+
- **Object Generalization** (libero_object)
|
| 48 |
+
- **Goal Generalization** (libero_goal)
|
| 49 |
+
- **Long-Horizon Tasks** (libero_100)
|
| 50 |
+
|
| 51 |
+
### Results
|
| 52 |
+
For detailed usage instructions, success rates, baseline comparisons, and evaluation protocols, please refer to the [OpenTau GitHub Repository](https://github.com/TensorAuto/OpenTau).
|