Robotics
PyTorch
robot-learning
world-action-model
flow-matching
libero
robocoin
galaxea
PoopBear commited on
Commit
dc1a082
·
verified ·
1 Parent(s): ae7c581

Simplify model card and link paper

Browse files
Files changed (1) hide show
  1. README.md +42 -90
README.md CHANGED
@@ -1,122 +1,74 @@
1
  ---
 
2
  license: apache-2.0
3
- library_name: fastwam
4
- pipeline_tag: robotics
5
  base_model: Wan-AI/Wan2.2-TI2V-5B
6
- base_model_relation: finetune
7
  datasets:
8
  - yuanty/LIBERO-fastwam
9
  tags:
10
  - robotics
 
11
  - world-action-model
12
  - flow-matching
13
  - libero
14
- - pytorch
15
  ---
16
 
17
- # RIFT
18
 
19
- This private model repository contains the canonical RIFT checkpoint trained
20
- on LIBERO. RIFT is derived from
21
- [FastWAM](https://github.com/yuantianyuan01/FastWAM) at commit
22
- `45d8e1458921d83f8ad6cf9ce993d371208dabd0`.
23
 
24
- RIFT uses one representation-shaping flow-matching head. The checkpoint also
25
- retains a detached L2 diagnostic probe for schema compatibility; this probe
26
- does not shape the shared anticipation representation.
27
 
28
- ## Files
 
 
29
 
30
- | File | Size | SHA-256 |
31
- |---|---:|---|
32
- | `rift_step021700.pt` | 12,060,942,777 bytes | `080e260b26c580cff5da0ce1de4c72e57eb2a6de47ae12f5712244dcf911f3cb` |
33
- | `dataset_stats.json` | 40,961 bytes | `f68a07ec322ea425164d94a8827aec1430109c7614637d03aec894aadc9fc05b` |
34
- | `config.yaml` | 2,370 bytes | `40620a22beb764177e2da1a235615a648692a0d3173f855d15e42b2084dc6e0c` |
35
 
36
- `dataset_stats.json` is the training-run normalization file paired with this
37
- checkpoint. Keep both files in the same directory for automatic discovery.
 
 
 
38
 
39
- ## Source and configuration
 
 
40
 
41
- - Source commit: [`184a582d7821e3a919dcd6e5795110ac0282eda6`](https://github.com/ChushanZhang/RIFT/tree/184a582d7821e3a919dcd6e5795110ac0282eda6)
42
- - Model configuration: [`config.yaml`](./config.yaml)
43
- - Canonical training task: [`configs/task/libero_rift_2cam224_1e-4.yaml`](https://github.com/ChushanZhang/RIFT/blob/184a582d7821e3a919dcd6e5795110ac0282eda6/configs/task/libero_rift_2cam224_1e-4.yaml)
44
- - Training step: 21,700
45
- - Checkpoint dtype: bfloat16
46
 
47
- The model depends on `Wan-AI/Wan2.2-TI2V-5B`, the tokenizer/text-encoder
48
- assets from `Wan-AI/Wan2.1-T2V-1.3B`, and the ActionDiT backbone generated by
49
- the preprocessing script in the source repository. The original training run
50
- did not record immutable Hub revisions for these external assets. Current
51
- dependency revisions must not be treated as proven training-time revisions.
52
 
53
  ## Download
54
 
55
  ```bash
56
- hf download PoopBear/RIFT \
57
- rift_step021700.pt dataset_stats.json config.yaml \
58
- --revision 9b03d72dbecb73e1cd805274ed26e7094412bb53 \
59
- --local-dir checkpoints/rift
60
- ```
61
-
62
- Because this repository is private, authenticate with `hf auth login` first.
63
-
64
- Verify the artifacts before loading:
65
-
66
- ```bash
67
- sha256sum checkpoints/rift/rift_step021700.pt \
68
- checkpoints/rift/dataset_stats.json \
69
- checkpoints/rift/config.yaml
70
- ```
71
-
72
- ## Install the matching source
73
 
74
- ```bash
75
- git clone https://github.com/ChushanZhang/RIFT.git
76
- cd RIFT
77
- git checkout 184a582d7821e3a919dcd6e5795110ac0282eda6
78
- pip install -e .
79
  ```
80
 
81
- The weight file contains `mot`, `proprio_encoder`, `step`, and `torch_dtype`.
82
- It is a weights-only checkpoint: optimizer, scheduler, and dataloader state are
83
- not included. It is released for evaluation only and is not a supported
84
- training warm start.
85
-
86
- ## Inputs and outputs
87
-
88
- The canonical LIBERO recipe uses two RGB cameras at 224 x 224, 8-dimensional
89
- proprioception, 7-dimensional actions, and a 32-step action horizon. Outputs
90
- must be interpreted using the included dataset statistics.
91
-
92
- ## Evaluation status
93
-
94
- This artifact release does not claim a new benchmark result. Before reporting
95
- results, pin the simulator commit, task suite, seeds, episodes per task,
96
- action-chunking policy, and aggregation method.
97
-
98
- ## Intended use and limitations
99
 
100
- The checkpoint is intended for research in controlled simulation. It is not
101
- intended for safety-critical control or unsupervised deployment on physical
102
- robots. It may fail under visual, linguistic, dynamics, or embodiment shift.
103
- Simulation performance does not establish real-world safety or robustness.
104
-
105
- PyTorch `.pt` files use pickle-based serialization. Load only the file whose
106
- checksum matches the value above, and prefer:
107
 
108
- ```python
109
- torch.load(path, map_location="cpu", mmap=True, weights_only=True)
 
 
 
 
 
110
  ```
111
 
112
- ## License and attribution
113
-
114
- The checkpoint is distributed under Apache License 2.0, consistent with the
115
- Wan base model license. RIFT source code is separately distributed under the
116
- license in the source repository. See [`NOTICE`](./NOTICE) for provenance and
117
- attribution.
118
-
119
- ## Citation
120
 
121
- Use the RIFT citation metadata from the matching source commit and also cite
122
- the upstream [Fast-WAM paper](https://arxiv.org/abs/2603.16666).
 
1
  ---
2
+ pretty_name: RIFT Checkpoint
3
  license: apache-2.0
4
+ library_name: pytorch
 
5
  base_model: Wan-AI/Wan2.2-TI2V-5B
 
6
  datasets:
7
  - yuanty/LIBERO-fastwam
8
  tags:
9
  - robotics
10
+ - robot-learning
11
  - world-action-model
12
  - flow-matching
13
  - libero
14
+ - arxiv:2608.11521
15
  ---
16
 
17
+ # RIFT Checkpoint
18
 
19
+ Official checkpoint for
20
+ **[Keep the Future, Drop the Rollout: RIFT for World Action Models](https://arxiv.org/abs/2608.11521)**.
 
 
21
 
22
+ ## Project
 
 
23
 
24
+ - Paper: https://arxiv.org/abs/2608.11521
25
+ - Code: https://github.com/ChushanZhang/RIFT
26
+ - Dataset: https://huggingface.co/datasets/yuanty/LIBERO-fastwam
27
 
28
+ ## Files
 
 
 
 
29
 
30
+ ```text
31
+ rift_step021700.pt
32
+ dataset_stats.json
33
+ config.yaml
34
+ ```
35
 
36
+ - `rift_step021700.pt`: released RIFT checkpoint for LIBERO
37
+ - `dataset_stats.json`: normalization statistics used by the checkpoint
38
+ - `config.yaml`: model configuration
39
 
40
+ Keep `dataset_stats.json` with the checkpoint during evaluation.
 
 
 
 
41
 
42
+ `rift_step021700.pt` uses PyTorch serialization. Download it only from this
43
+ repository and load it with `weights_only=True`.
 
 
 
44
 
45
  ## Download
46
 
47
  ```bash
48
+ pip install -U huggingface_hub
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
49
 
50
+ hf download PoopBear/RIFT \
51
+ rift_step021700.pt \
52
+ dataset_stats.json \
53
+ config.yaml \
54
+ --local-dir ./checkpoints/rift
55
  ```
56
 
57
+ Evaluation commands are provided in the
58
+ [RIFT repository](https://github.com/ChushanZhang/RIFT#benchmark-evaluation).
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
59
 
60
+ ## Citation
 
 
 
 
 
 
61
 
62
+ ```bibtex
63
+ @article{zhang2026rift,
64
+ title={Keep the Future, Drop the Rollout: RIFT for World Action Models},
65
+ author={Zhang, Chushan and Tong, Jinguang and Li, Xuesong and Wang, Yikai and Li, Hongdong},
66
+ journal={arXiv preprint arXiv:2608.11521},
67
+ year={2026}
68
+ }
69
  ```
70
 
71
+ ## License
 
 
 
 
 
 
 
72
 
73
+ The checkpoint is distributed under the Apache License 2.0. RIFT source code
74
+ is distributed separately under the MIT License. See `NOTICE` for attribution.