LIBERO: release the H16 model reported in the paper (16-step chunks, replan 16)

libero/config.yaml

@@ -9,7 +9,7 @@ model:
   rgb_input_size:
   - 224
   - 448
-  rgb_target_mode: f8
+  rgb_target_mode: f4
   rgb_depth_attention_mode: rgb_to_depth_to_action
   rgb_fast: true
   rgb_history_num_frames: 5
@@ -102,17 +102,17 @@ data:
     video_size:
     - 224
     - 448
-    num_video_frames: 9
+    num_video_frames: 5
     frame_stride: 4
     rgb_history_num_frames: 5
     rgb_history_frame_stride: 4
     cameras:
     - agentview
     - eye_in_hand
-    action_horizon: 32
+    action_horizon: 16
     context_len: 128
     text_encoder_id: wan22ti2v5b
-    rgb_target_mode: f8
+    rgb_target_mode: f4
     depth_max_meters: 1.5
     processor:
       _target_: prismwam.datasets.lerobot.processors.prismwam_processor.PrismWAMProcessor
@@ -144,7 +144,7 @@ data:
         - key: default
           raw_shape: 8
           shape: 8
-      num_obs_steps: 33
+      num_obs_steps: 17
       num_output_cameras: 2
       action_output_dim: 7
       proprio_output_dim: 8
@@ -160,7 +160,7 @@ data:
       action_state_merger:
         _target_: prismwam.datasets.lerobot.transforms.action_state_merger.ConcatLeftAlign
 deployment:
-  action_horizon: 32
+  action_horizon: 16
   replan_steps: 16
   num_inference_steps: 10
   inference_shifts:

libero/policy.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d88b7bd35538d6c325ac66c24f00bdb62a5ae0e8227b5aa190065e444134cdb9
+oid sha256:b7fec1e77142f9d80846dc5aa760805eab16d85185886ad7784d886c6052acd7
 size 12222027990