1, add input normalization in configuration_sac.py 2, add masking on loss computation

lerobot/common/policies/sac/configuration_sac.py

@@ -28,12 +28,18 @@ class SACConfig:
     )
     output_shapes: dict[str, list[int]] = field(
         default_factory=lambda: {
-            "action": [4],
+            "action": [2],
         }
     )
 
     # Normalization / Unnormalization
-    input_normalization_modes: dict[str, str] | None = None
+    input_normalization_modes: dict[str, str] = field(
+        default_factory=lambda: {
+            "observation.image": "mean_std",
+            "observation.state": "min_max",
+            "observation.environment_state": "min_max",
+        }
+    )
     output_normalization_modes: dict[str, str] = field(
         default_factory=lambda: {"action": "min_max"},
     )

lerobot/common/policies/sac/modeling_sac.py

@@ -177,25 +177,19 @@ class SACPolicy(
 
         # 4- Calculate loss
         # Compute state-action value loss (TD loss) for all of the Q functions in the ensemble.
-        critics_loss = F.mse_loss(
+        critics_loss = (   
-            q_preds,  # shape: [num_critics, batch_size]
+            F.mse_loss(
-            einops.repeat(td_target, "b -> e b", e=q_preds.shape[0]), # expand td_target to match q_preds shape
+                    q_preds,
-            reduction="none"
+                    einops.repeat(td_target, "b -> e b", e=q_preds.shape[0]),
-        ).sum(0).mean()
+                    reduction="none",
-
+                ).sum(0)  # sum over ensemble
-        # critics_loss = (   
+                # `q_preds_ensemble` depends on the first observation and the actions.
-        #     F.mse_loss(
+                * ~batch["observation.state_is_pad"][:,0] # shape: [batch_size, horizon+1]
-        #             q_preds,
+                * ~batch["action_is_pad"][:,0] # shape: [batch_size, horizon]
-        #             einops.repeat(td_target, "b -> e b", e=q_preds.shape[0]),
+                # q_targets depends on the reward and the next observations.
-        #             reduction="none",
+                * ~batch["next.reward_is_pad"][:,0] # shape: [batch_size, horizon]
-        #         ).sum(0)  # sum over ensemble
+                * ~batch["observation.state_is_pad"][:,1] # shape: [batch_size, horizon+1]
-        #         # `q_preds_ensemble` depends on the first observation and the actions.
+            ).mean()
-        #         * ~batch["observation.state_is_pad"][0]
-        #         * ~batch["action_is_pad"]
-        #         # q_targets depends on the reward and the next observations.
-        #         * ~batch["next.reward_is_pad"]
-        #         * ~batch["observation.state_is_pad"][1:]
-        #     ).sum(0).mean()
         
         # calculate actors loss
         # 1- temperature
@@ -207,8 +201,8 @@ class SACPolicy(
             q_preds = self.critic_forward(observations, actions, use_target=False)
         actor_loss = (
             -(q_preds - temperature * log_probs).mean()
-            # * ~batch["observation.state_is_pad"][0]
+            * ~batch["observation.state_is_pad"][:,0] # shape: [batch_size, horizon+1]
-            # * ~batch["action_is_pad"]
+            * ~batch["action_is_pad"][:,0] # shape: [batch_size, horizon]
         ).mean()
 
 

lerobot/configs/policy/sac_pusht_keypoints.yaml

@@ -19,7 +19,7 @@ training:
   grad_clip_norm: 10.0
   lr: 3e-4
 
-  eval_freq: 50000
+  eval_freq: 10000
   log_freq: 500
   save_freq: 50000
 
@@ -46,8 +46,8 @@ policy:
 
   # Input / output structure.
   n_action_repeats: 1
-  horizon: 5
+  horizon: 2
-  n_action_steps: 5
+  n_action_steps: 2
 
   input_shapes:
     # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?