Robotics
PyTorch
robot-learning
world-action-model
flow-matching
libero
robocoin
galaxea
PoopBear commited on
Commit
b10a6f1
·
verified ·
1 Parent(s): dc1a082

Update RIFT package namespace

Browse files

Sync the published Hydra target and model card with the standalone rift package.

Files changed (2) hide show
  1. README.md +6 -1
  2. config.yaml +1 -8
README.md CHANGED
@@ -35,10 +35,15 @@ config.yaml
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
 
 
35
 
36
  - `rift_step021700.pt`: released RIFT checkpoint for LIBERO
37
  - `dataset_stats.json`: normalization statistics used by the checkpoint
38
+ - `config.yaml`: current Hydra model configuration using the `rift.*` namespace
39
 
40
  Keep `dataset_stats.json` with the checkpoint during evaluation.
41
 
42
+ Install the standalone [RIFT repository](https://github.com/ChushanZhang/RIFT)
43
+ with `pip install -e .`. Compose `config.yaml` with the repository's task and
44
+ data configs; it contains interpolations such as `${data.train...}` and is not
45
+ a complete training configuration by itself.
46
+
47
  `rift_step021700.pt` uses PyTorch serialization. Download it only from this
48
  repository and load it with `weights_only=True`.
49
 
config.yaml CHANGED
@@ -1,27 +1,20 @@
1
- _target_: fastwam.rift.create_fastwam_anticip
2
 
3
- # Canonical RIFT: full anticipation grid, two-pass video co-training,
4
- # flow-matching render objective, and a detached L2 diagnostic probe retained
5
- # for checkpoint-schema compatibility.
6
  anticip_future_tokens: 196
7
  anticip_latent_t: 3
8
  anticip_lambda: 1.0
9
  anticip_action_lambda: 1.0
10
 
11
- # Motion-aware target and late loss rebalance.
12
  anticip_delta_mix: 0.8
13
  anticip_lambda_final: 0.2
14
  anticip_anneal_start_frac: 0.7
15
  anticip_anneal_total_steps: 21700
16
 
17
- # Late render-only conditioning-noise curriculum.
18
  anticip_cond_noise_p: 0.3
19
  anticip_cond_noise_sigma: 0.06
20
  anticip_cond_noise_ramp_start_frac: 0.7
21
  anticip_cond_noise_action_free: true
22
 
23
- # The FM head is the released representation-shaping head. The L2 probe
24
- # remains a detached diagnostic probe and does not shape anticipation tokens.
25
  anticip_fm_lambda: 1.0
26
  anticip_fm_width: 512
27
  anticip_fm_blocks: 2
 
1
+ _target_: rift.model_factory.create_rift
2
 
 
 
 
3
  anticip_future_tokens: 196
4
  anticip_latent_t: 3
5
  anticip_lambda: 1.0
6
  anticip_action_lambda: 1.0
7
 
 
8
  anticip_delta_mix: 0.8
9
  anticip_lambda_final: 0.2
10
  anticip_anneal_start_frac: 0.7
11
  anticip_anneal_total_steps: 21700
12
 
 
13
  anticip_cond_noise_p: 0.3
14
  anticip_cond_noise_sigma: 0.06
15
  anticip_cond_noise_ramp_start_frac: 0.7
16
  anticip_cond_noise_action_free: true
17
 
 
 
18
  anticip_fm_lambda: 1.0
19
  anticip_fm_width: 512
20
  anticip_fm_blocks: 2