online training works (loss goes down), remove repeat_action, eval_policy outputs episodes data, eval_policy uses max_episodes_rendered

lerobot/configs/env/aloha.yaml

@@ -18,7 +18,6 @@ env:
   from_pixels: True
   pixels_only: False
   image_size: [3, 480, 640]
-  action_repeat: 1
   episode_length: 400
   fps: ${fps}
 

lerobot/configs/env/pusht.yaml

@@ -18,7 +18,6 @@ env:
   from_pixels: True
   pixels_only: False
   image_size: 96
-  action_repeat: 1
   episode_length: 300
   fps: ${fps}
 

lerobot/configs/env/xarm.yaml

@@ -17,7 +17,6 @@ env:
   from_pixels: True
   pixels_only: False
   image_size: 84
-  # action_repeat: 2  # we can remove if policy has n_action_steps=2
   episode_length: 25
   fps: ${fps}
 

lerobot/configs/policy/tdmpc.yaml

@@ -36,6 +36,7 @@ policy:
   log_std_max: 2
 
   # learning
+  batch_size: 256
   max_buffer_size: 10000
   horizon: 5
   reward_coef: 0.5