Create README
Browse files
README.md
ADDED
@@ -0,0 +1,133 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
'[object Object]': null
|
3 |
+
license: mit
|
4 |
+
language:
|
5 |
+
- en
|
6 |
+
---
|
7 |
+
|
8 |
+
# Model Card for 3D Diffuser Actor
|
9 |
+
|
10 |
+
<!-- Provide a quick summary of what the model is/does. -->
|
11 |
+
|
12 |
+
A robot manipulation policy that marries diffusion modeling with 3D scene representations.
|
13 |
+
3D Diffuser Actor is trained and evaluated on [RLBench](https://github.com/stepjam/RLBench) or [CALVIN](https://github.com/mees/calvin) simulation.
|
14 |
+
We release all code, checkpoints, and details involved in training these models.
|
15 |
+
|
16 |
+
## Model Details
|
17 |
+
|
18 |
+
The models released are the following:
|
19 |
+
|
20 |
+
| Benchmark | Embedding dimension | Diffusion timestep |
|
21 |
+
|------|------|------|
|
22 |
+
| [RLBench (PerAct)]() | 120 | 100 |
|
23 |
+
| [RLBench (GNFactor)]() | 120| 100 |
|
24 |
+
| [CALVIN]() | 192 | 25 |
|
25 |
+
|
26 |
+
### Model Description
|
27 |
+
|
28 |
+
<!-- Provide a longer summary of what this model is. -->
|
29 |
+
|
30 |
+
|
31 |
+
- **Developed by:** Katerina Group at CMU
|
32 |
+
- **Model type:** a Diffusion model with 3D scene
|
33 |
+
- **License:** The code and model are released under MIT License
|
34 |
+
- **Contact:** ngkanats@andrew.cmu.edu
|
35 |
+
|
36 |
+
|
37 |
+
### Model Sources [optional]
|
38 |
+
|
39 |
+
<!-- Provide the basic links for the model. -->
|
40 |
+
|
41 |
+
- **Project Page:** https://3d-diffuser-actor.github.io
|
42 |
+
- **Repository:** https://github.com/nickgkan/3d_diffuser_actor.git
|
43 |
+
- **Paper:** [Link]()
|
44 |
+
|
45 |
+
## Uses
|
46 |
+
|
47 |
+
<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
|
48 |
+
|
49 |
+
TODO
|
50 |
+
|
51 |
+
### Direct Use
|
52 |
+
|
53 |
+
<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
|
54 |
+
|
55 |
+
TODO
|
56 |
+
|
57 |
+
|
58 |
+
## Evaluation
|
59 |
+
|
60 |
+
<!-- This section describes the evaluation protocols and provides the results. -->
|
61 |
+
Our model trained and evaluated on RLBench simulation with the PerAct setup:
|
62 |
+
|
63 |
+
| RLBench (PerAct) | 3D Diffuser Actor | [RVT](https://github.com/NVlabs/RVT) |
|
64 |
+
| --------------------------------- | -------- | -------- |
|
65 |
+
| average | 81.3 | 62.9 |
|
66 |
+
| open drawer | 89.6 | 71.2 |
|
67 |
+
| slide block | 97.6 | 81.6 |
|
68 |
+
| sweep to dustpan | 84.0 | 72.0 |
|
69 |
+
| meat off grill | 96.8 | 88 |
|
70 |
+
| turn tap | 99.2 | 93.6 |
|
71 |
+
| put in drawer | 96.0 | 88.0 |
|
72 |
+
| close jar | 96.0 | 52.0 |
|
73 |
+
| drag stick | 100.0 | 99.2 |
|
74 |
+
| stack blocks | 68.3 | 28.8 |
|
75 |
+
| screw bulbs | 82.4 | 48.0 |
|
76 |
+
| put in safe | 97.6 | 91.2 |
|
77 |
+
| place wine | 93.6 | 91.0 |
|
78 |
+
| put in cupboard | 85.6 | 49.6 |
|
79 |
+
| sort shape | 44.0 | 36.0 |
|
80 |
+
| push buttons | 98.4 | 100.0 |
|
81 |
+
| insert peg | 65.6 | 11.2 |
|
82 |
+
| stack cups | 47.2 | 26.4 |
|
83 |
+
| place cups | 24.0 | 4.0 |
|
84 |
+
|
85 |
+
|
86 |
+
Our model trained and evaluated on RLBench simulation with the GNFactor setup:
|
87 |
+
|
88 |
+
| RLBench (PerAct) | 3D Diffuser Actor | [GNFactor](https://github.com/YanjieZe/GNFactor) |
|
89 |
+
| --------------------------------- | -------- | -------- |
|
90 |
+
| average | 78.4 | 31.7 |
|
91 |
+
| open drawer | 89.3 | 76.0 |
|
92 |
+
| sweep to dustpan | 894.7 | 25.0 |
|
93 |
+
| close jar | 82.7 | 25.3 |
|
94 |
+
| meat off grill | 88.0 | 57.3 |
|
95 |
+
| turn tap | 80.0 | 50.7 |
|
96 |
+
| slide block | 92.0 | 20.0 |
|
97 |
+
| put in drawer | 77.3 | 0.0 |
|
98 |
+
| drag stick | 98.7 | 37.3 |
|
99 |
+
| push buttons | 69.3 | 18.7 |
|
100 |
+
| stack blocks | 12.0 | 4.0 |
|
101 |
+
|
102 |
+
Our model trained and evaluated on CALVIN simulation (train with environment A, B, C and test on D):
|
103 |
+
|
104 |
+
| RLBench (PerAct) | 3D Diffuser Actor | [GR-1](https://gr1-manipulation.github.io/) | [SuSIE](https://rail-berkeley.github.io/susie/) |
|
105 |
+
| --------------------------------- | -------- | -------- | -------- |
|
106 |
+
| task 1 | 92.2 | 85.4 | 87.0 |
|
107 |
+
| task 2 | 78.7 | 71.2 | 69.0 |
|
108 |
+
| task 3 | 63.9 | 59.6 | 49.0 |
|
109 |
+
| task 4 | 51.2 | 49.7 | 38.0 |
|
110 |
+
| task 5 | 41.2 | 40.1 | 26.0 |
|
111 |
+
|
112 |
+
|
113 |
+
|
114 |
+
## Citation [optional]
|
115 |
+
|
116 |
+
<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
|
117 |
+
|
118 |
+
**BibTeX:**
|
119 |
+
|
120 |
+
```
|
121 |
+
@article{,
|
122 |
+
title={Action Diffusion with 3D Scene Representations},
|
123 |
+
author={Ke, Tsung-Wei and Gkanatsios, Nikolaos and Fragkiadaki, Katerina}
|
124 |
+
journal={Preprint},
|
125 |
+
year={2024}
|
126 |
+
}
|
127 |
+
```
|
128 |
+
|
129 |
+
|
130 |
+
|
131 |
+
## Model Card Contact
|
132 |
+
|
133 |
+
For errors in this model card, contact Nikos or Tsung-Wei, {ngkanats, tsungwek} at andrew dot cmu dot edu.
|