Upload policy weights, train config and readme

README.md

@@ -5,8 +5,8 @@ license: apache-2.0
 model_name: lawam
 pipeline_tag: robotics
 tags:
-- robotics
 - lawam
+- robotics
 - lerobot
 ---
 
@@ -83,7 +83,7 @@ The policy consumes these observation features and produces these action feature
 | Optimizer | adamw |
 | Learning rate | 0.0001 |
 | Seed | 1000 |
-| LeRobot version | 0.6.1 |
+| LeRobot version | 0.6.2 |
 
 ---
 

config.json

@@ -52,9 +52,39 @@
         "ACTION": "MIN_MAX"
     },
     "base_vlm": "Qwen/Qwen3-VL-2B-Instruct",
-    "lam_ckpt_path": "/fsx/maxime_ellerbach/ckpts/lawam_lam/checkpoints/pytorch_model.pt",
-    "lam_yaml_path": "/fsx/maxime_ellerbach/robots/runs/lawam/review_pr3999/dino_large_vae_localdinov3.yaml",
+    "base_vlm_path": null,
+    "lam_ckpt_path": null,
     "hf_cache_dir": null,
+    "lam_dim": 1024,
+    "lam_num_heads": 16,
+    "lam_ffn_expansion_factor": 4,
+    "lam_enc_layers": 24,
+    "lam_codebook_size": 32,
+    "lam_code_dim": 32,
+    "lam_max_state_dim": 14,
+    "lam_num_queries": 1,
+    "lam_dec_layers": 12,
+    "lam_dropout": 0.0,
+    "lam_vq_type": "vae",
+    "lam_vq_layer_norm": true,
+    "lam_norm_latents": true,
+    "lam_norm_latents_type": "ln",
+    "lam_enc_add_state": false,
+    "lam_enc_modal_mask": true,
+    "lam_latent_layer_to_use": -2,
+    "lam_multi_input": false,
+    "lam_num_embodiments": 32,
+    "lam_image_hw": [
+        256,
+        256
+    ],
+    "lam_patch_size": 16,
+    "lam_decoder_last_ln": true,
+    "dinov3_hidden_size": 768,
+    "dinov3_intermediate_size": 3072,
+    "dinov3_num_hidden_layers": 12,
+    "dinov3_num_attention_heads": 12,
+    "dinov3_num_register_tokens": 4,
     "lawam_checkpoint_path": null,
     "lawam_dataset_stats_path": null,
     "lawam_unnorm_key": null,
@@ -104,11 +134,18 @@
     "flow_noise_s": 0.999,
     "flow_token_independent_noise": false,
     "flow_use_action_positional_embeddings": true,
-    "clip_normalized_actions": true,
-    "pre_snap_gripper_action": true,
-    "binarize_gripper_action": true,
+    "clip_normalized_actions": false,
+    "pre_snap_gripper_action": false,
+    "binarize_gripper_action": false,
     "gripper_dim": 6,
     "gripper_threshold": 0.5,
+    "freeze_vision_backbone": false,
+    "freeze_llm_backbone": false,
+    "freeze_embedding": true,
+    "unfreeze_vision_merger": true,
+    "unfreeze_lam_decoder": true,
+    "keep_llm_first_n_layers": 16,
+    "unfreeze_llm_last_n_layers": -1,
     "optimizer_lr": 0.0001,
     "optimizer_betas": [
         0.9,

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f1ca6d1a5be545eb062d0be82d34c91e56fe6faaf1763afbd7c21cd61a2cb8cb
-size 8382100720
+oid sha256:f6cb1481708e3e33ad0131f407153be943260c6bc41ab09041834938c7c3e9b1
+size 7173993832

train_config.json

@@ -1,8 +1,10 @@
 {
     "dataset": {
         "repo_id": "maximellerbach/omx_multicubes",
+        "repo_type": "dataset",
         "root": null,
         "episodes": null,
+        "exclude_episodes": null,
         "image_transforms": {
             "enable": false,
             "max_num_transforms": 3,
@@ -137,9 +139,39 @@
             "ACTION": "MIN_MAX"
         },
         "base_vlm": "Qwen/Qwen3-VL-2B-Instruct",
+        "base_vlm_path": null,
         "lam_ckpt_path": "/fsx/maxime_ellerbach/ckpts/lawam_lam/checkpoints/pytorch_model.pt",
-        "lam_yaml_path": "/fsx/maxime_ellerbach/robots/runs/lawam/review_pr3999/dino_large_vae_localdinov3.yaml",
         "hf_cache_dir": null,
+        "lam_dim": 1024,
+        "lam_num_heads": 16,
+        "lam_ffn_expansion_factor": 4,
+        "lam_enc_layers": 24,
+        "lam_codebook_size": 32,
+        "lam_code_dim": 32,
+        "lam_max_state_dim": 14,
+        "lam_num_queries": 1,
+        "lam_dec_layers": 12,
+        "lam_dropout": 0.0,
+        "lam_vq_type": "vae",
+        "lam_vq_layer_norm": true,
+        "lam_norm_latents": true,
+        "lam_norm_latents_type": "ln",
+        "lam_enc_add_state": false,
+        "lam_enc_modal_mask": true,
+        "lam_latent_layer_to_use": -2,
+        "lam_multi_input": false,
+        "lam_num_embodiments": 32,
+        "lam_image_hw": [
+            256,
+            256
+        ],
+        "lam_patch_size": 16,
+        "lam_decoder_last_ln": true,
+        "dinov3_hidden_size": 768,
+        "dinov3_intermediate_size": 3072,
+        "dinov3_num_hidden_layers": 12,
+        "dinov3_num_attention_heads": 12,
+        "dinov3_num_register_tokens": 4,
         "lawam_checkpoint_path": null,
         "lawam_dataset_stats_path": null,
         "lawam_unnorm_key": null,
@@ -189,11 +221,18 @@
         "flow_noise_s": 0.999,
         "flow_token_independent_noise": false,
         "flow_use_action_positional_embeddings": true,
-        "clip_normalized_actions": true,
-        "pre_snap_gripper_action": true,
-        "binarize_gripper_action": true,
+        "clip_normalized_actions": false,
+        "pre_snap_gripper_action": false,
+        "binarize_gripper_action": false,
         "gripper_dim": 6,
         "gripper_threshold": 0.5,
+        "freeze_vision_backbone": false,
+        "freeze_llm_backbone": false,
+        "freeze_embedding": true,
+        "unfreeze_vision_merger": true,
+        "unfreeze_lam_decoder": true,
+        "keep_llm_first_n_layers": 16,
+        "unfreeze_llm_last_n_layers": -1,
         "optimizer_lr": 0.0001,
         "optimizer_betas": [
             0.9,