Hugging Face's logo

armteam
/
crab-smolvla-rwfm

Robotics
manipulation
smolvla
vision-language-action
tactile
reward-weighted-flow-matching
Model card Files
xet
 Community
businesslion commited on
Commit
7082087
·
verified ·
1 Parent(s): 993974f

Add training config

Browse files
Files changed (1) hide show
  1. config.yaml +148 -0
config.yaml ADDED
@@ -0,0 +1,148 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # ============================================================================
2
+ # RWFM right arm 6-DOF + tactile — v3 with data augmentation + improved hyperparams.
3
+ # Uses vla_2.0 labeled datasets with per-step reward column + episode_reward.
4
+ # All 34 datasets: 27 clean + 7 crack, reward weighting active.
5
+ # Action/state: right arm only [6-11].
6
+ # Changes from v2: augmentation, weight_decay=0.01, grad_accum=4 (eff batch=32).
7
+ # ============================================================================
8
+
9
+ experiment:
10
+ name: "crab_smolvla_6dof_right_arm_rwfm_v3"
11
+ seed: 42
12
+ output_dir: "outputs/crab_smolvla_6dof_right_arm_rwfm_v3"
13
+ resume_from_checkpoint: null
14
+
15
+ model:
16
+ pretrained_path: "lerobot/smolvla_base"
17
+ chunk_size: 50
18
+ n_action_steps: 50
19
+ max_state_dim: 32
20
+ max_action_dim: 32
21
+ action_dim: 6
22
+ state_dim: 6
23
+ num_steps: 10
24
+ freeze_vision_encoder: true
25
+ train_expert_only: true
26
+ train_state_proj: true
27
+ tactile_encoder:
28
+ enabled: true
29
+ input_dim: 100
30
+ hidden_dim: 128
31
+ output_dim: 64
32
+ num_layers: 2
33
+ activation: "gelu"
34
+ dropout: 0.1
35
+
36
+ dataset:
37
+ episodes_dir: "../datasets_labeled"
38
+ episode_names:
39
+ # --- Egg 12V clean (11 datasets) ---
40
+ - "egg_carton_to_tray/pick_place_soft_20260217_1847"
41
+ - "egg_carton_to_tray/pick_place_soft_20260217_1905"
42
+ - "egg_carton_to_tray/pick_place_soft_20260217_2003"
43
+ - "egg_carton_to_tray/pick_place_soft_20260217_2020"
44
+ - "egg_carton_to_tray/pick_place_soft_20260218_1414"
45
+ - "egg_carton_to_tray/pick_place_soft_20260218_1422"
46
+ - "egg_carton_to_tray/pick_place_soft_20260218_1429"
47
+ - "egg_carton_to_tray/pick_place_soft_20260218_1446"
48
+ - "egg_carton_to_tray/pick_place_soft_20260219_1759"
49
+ - "egg_carton_to_tray/pick_place_medium_20260226_1638"
50
+ - "egg_carton_to_tray/pick_place_medium_20260226_1649"
51
+ # --- Marmalade jar clean (7 datasets) ---
52
+ - "marmalade_jar/pick_place_medium_20260226_1659"
53
+ - "marmalade_jar/pick_place_medium_20260226_1707"
54
+ - "marmalade_jar/pick_place_medium_20260226_1713_2"
55
+ - "marmalade_jar/pick_place_medium_20260226_1717"
56
+ - "marmalade_jar/pick_place_medium_20260226_1722"
57
+ - "marmalade_jar/pick_place_medium_20260227_1636"
58
+ - "marmalade_jar/pick_place_medium_20260227_1641"
59
+ # --- Waffles clean (9 datasets) ---
60
+ - "waffles/pick_place_soft_20260226_1727"
61
+ - "waffles/pick_place_soft_20260226_1734"
62
+ - "waffles/pick_place_soft_20260226_1738"
63
+ - "waffles/pick_place_soft_20260226_1743"
64
+ - "waffles/pick_place_soft_20260226_1748"
65
+ - "waffles/pick_place_soft_20260227_1507"
66
+ - "waffles/pick_place_soft_20260227_1611"
67
+ - "waffles/pick_place_soft_20260227_1616"
68
+ - "waffles/pick_place_soft_20260227_1621"
69
+ # --- Egg cracked (3 datasets, negative reward) ---
70
+ - "egg_carton_to_tray_cracked/pick_place_soft_20260219_1820"
71
+ - "egg_carton_to_tray_cracked/pick_place_soft_20260219_1831"
72
+ - "egg_carton_to_tray_cracked/pick_place_soft_20260219_1839"
73
+ # --- Marmalade crack (2 datasets, negative reward) ---
74
+ - "marmalade_jar_crack/pick_place_medium_20260227_1643"
75
+ - "marmalade_jar_crack/pick_place_medium_20260227_1648"
76
+ # --- Waffles crack (2 datasets, negative reward) ---
77
+ - "waffles_crack/pick_place_soft_20260227_1627"
78
+ - "waffles_crack/pick_place_soft_20260227_1633"
79
+ action_indices: [6, 7, 8, 9, 10, 11]
80
+ state_indices: [6, 7, 8, 9, 10, 11]
81
+ image_keys:
82
+ - "observation.images.main_camera"
83
+ - "observation.images.left_arm_camera"
84
+ - "observation.images.right_arm_camera"
85
+ fps: 15
86
+ image_size: [256, 256]
87
+ original_image_size: [480, 640]
88
+ video_codec: "av1"
89
+ task_description: "Pick and place the object"
90
+ val_ratio: 0.1
91
+ num_workers: 0
92
+ video_cache_dir: "/tmp/crab_video_cache_labeled"
93
+ augmentation:
94
+ enabled: true
95
+ color_jitter:
96
+ brightness: 0.2
97
+ contrast: 0.2
98
+ saturation: 0.2
99
+ hue: 0.05
100
+ random_crop:
101
+ enabled: true
102
+ scale: [0.9, 1.0]
103
+
104
+ training:
105
+ batch_size: 8
106
+ gradient_accumulation_steps: 4
107
+ steps: 50000
108
+ warmup_steps: 2000
109
+ optimizer: "adamw"
110
+ learning_rate: 1.0e-4
111
+ weight_decay: 0.01
112
+ betas: [0.9, 0.95]
113
+ eps: 1.0e-8
114
+ grad_clip_norm: 10.0
115
+ scheduler: "cosine_with_warmup"
116
+ decay_steps: 50000
117
+ min_lr: 2.5e-6
118
+ use_amp: true
119
+ amp_dtype: "bfloat16"
120
+ checkpoint:
121
+ save_every_steps: 5000
122
+ keep_last_n: 5
123
+ save_optimizer: true
124
+ save_best: true
125
+ log_every_steps: 50
126
+ eval_every_steps: 2500
127
+ visualize_every_steps: 1000
128
+
129
+ reward_weighting:
130
+ enabled: true
131
+ alpha: 0.25
132
+ alpha_warmup_steps: 1000
133
+ alpha_ramp_steps: 3000
134
+ w_min: 0.25
135
+ w_max: 4.0
136
+ mix_beta: 0.7
137
+ z_clip: 6.0
138
+ a_clip: 6.0
139
+ group_key: "task"
140
+ eps: 1.0e-6
141
+
142
+ anchor:
143
+ enabled: true
144
+ lambda: 1.0e-4
145
+ warmup_steps: 2000
146
+ ramp_steps: 6000
147
+ trainable_only: true
148
+ reference_device: "cpu"