File size: 1,957 Bytes
4844da9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
22f3b57
4844da9
22f3b57
4844da9
 
 
 
 
 
 
 
 
 
 
 
dbc69a4
4844da9
 
 
 
bcda19c
4844da9
 
 
 
 
 
 
 
 
 
bcda19c
4844da9
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
---
library_name: opentau
tags:
- robotics
- vla
- pi05
- libero
- manipulation
- flow-matching
- pytorch
base_model: williamyue/pi05_base
license: apache-2.0
datasets:
- physical-intelligence/libero
repo_url: https://github.com/TensorAuto/OpenTau
---

# tPi0.5-libero

A **pi0.5 (π₀.₅)** Vision-Language-Action (VLA) model, finetuned on the **LIBERO** robotic manipulation benchmark using the **OpenTau** training framework. This model is designed to follow natural language instructions to perform manipulation tasks in a simulated tabletop environment.

**For full documentation, evaluation results, and inference code, please visit the repository:**
<br>
👉 **[https://github.com/TensorAuto/OpenTau](https://github.com/TensorAuto/OpenTau)**

---

## Model Details

### Description
- **Model Type:** Vision-Language-Action (VLA) Model
- **Base Architecture:** π₀.₅ (pi0.5) by Physical Intelligence
- **Backbone:** PaliGemma-3B (VLM) + Gemma-300M (Action Expert)
- **Training Data:** LIBERO (Lifelong Robot Learning) Benchmark
- **Framework:** OpenTau

### Architecture
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 (10-step action chunks) via flow matching.

---

## Training and Evaluation

### Dataset
This model was finetuned on the **LIBERO** benchmark dataset. The LIBERO suite consists of human-teleoperated demonstrations for tabletop manipulation, covering:
- **Spatial Generalization** (libero_spatial)
- **Object Generalization** (libero_object)
- **Goal Generalization** (libero_goal)
- **Long-Horizon Tasks** (libero_10)

### Results
For detailed usage instructions, success rates, baseline comparisons, and evaluation protocols, please refer to the [OpenTau GitHub Repository](https://github.com/TensorAuto/OpenTau).