fix(train): enable relative action overrides for pretrained processors (#3711)

* fix(train): enable relative action overrides for pretrained processors
Keep pretrained processor pipelines when use_relative_actions is enabled and
apply relative/absolute action processor settings through overrides. Rename the
relative action processor registry key to relative_actions_processor.

* fix(config): reject rename_map without pretrained checkpoint

Fail fast when rename_map is set during fresh initialization, since fresh
configs derive feature names from the current dataset and no rename is applied.

---------

Co-authored-by: Pepijn <138571049+pkooij@users.noreply.github.com>

src/lerobot/configs/train.py

@@ -177,6 +177,12 @@ class TrainPipelineConfig(HubMixin):
             )
 
         active_cfg = self.trainable_config
+        if self.rename_map and active_cfg.pretrained_path is None:
+            raise ValueError(
+                "`rename_map` requires a pretrained policy checkpoint. "
+                "Fresh initialization derives feature names from the current dataset, so no rename is applied."
+            )
+
         if not self.job_name:
             if self.env is None:
                 self.job_name = f"{active_cfg.type}"

src/lerobot/processor/relative_action_processor.py

@@ -81,7 +81,7 @@ def to_absolute_actions(actions: Tensor, state: Tensor, mask: Sequence[bool]) ->
     return actions
 
 
-@ProcessorStepRegistry.register("delta_actions_processor")
+@ProcessorStepRegistry.register("relative_actions_processor")
 @dataclass
 class RelativeActionsProcessorStep(ProcessorStep):
     """Converts absolute actions to relative actions (action -= state) for masked dimensions.

src/lerobot/scripts/lerobot_train.py

@@ -292,19 +292,8 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
 
     active_cfg = cfg.trainable_config
     processor_pretrained_path = active_cfg.pretrained_path
-    if (
-        getattr(active_cfg, "use_relative_actions", False)
-        and processor_pretrained_path is not None
-        and not cfg.resume
-    ):
-        logging.warning(
-            "use_relative_actions=true with pretrained processors can skip relative transforms if "
-            "the checkpoint processors do not define them. Building processors from current policy config."
-        )
-        processor_pretrained_path = None
 
     processor_kwargs = {}
-    postprocessor_kwargs = {}
     if (processor_pretrained_path and not cfg.resume) or not processor_pretrained_path:
         processor_kwargs["dataset_stats"] = dataset.meta.stats
 
@@ -312,24 +301,31 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
         processor_kwargs["dataset_meta"] = dataset.meta
 
     if not cfg.is_reward_model_training and processor_pretrained_path is not None:
-        processor_kwargs["preprocessor_overrides"] = {
+        preprocessor_overrides = {
             "device_processor": {"device": device.type},
             "normalizer_processor": {
                 "stats": dataset.meta.stats,
                 "features": {**policy.config.input_features, **policy.config.output_features},
                 "norm_map": policy.config.normalization_mapping,
             },
+            "rename_observations_processor": {"rename_map": cfg.rename_map},
         }
-        processor_kwargs["preprocessor_overrides"]["rename_observations_processor"] = {
+        postprocessor_overrides = {
-            "rename_map": cfg.rename_map
-        }
-        postprocessor_kwargs["postprocessor_overrides"] = {
             "unnormalizer_processor": {
                 "stats": dataset.meta.stats,
                 "features": policy.config.output_features,
                 "norm_map": policy.config.normalization_mapping,
             },
         }
+        if getattr(active_cfg, "use_relative_actions", False):
+            preprocessor_overrides["relative_actions_processor"] = {
+                "enabled": True,
+                "exclude_joints": getattr(active_cfg, "relative_exclude_joints", []),
+                "action_names": getattr(active_cfg, "action_feature_names", None),
+            }
+            postprocessor_overrides["absolute_actions_processor"] = {"enabled": True}
+        processor_kwargs["preprocessor_overrides"] = preprocessor_overrides
+        processor_kwargs["postprocessor_overrides"] = postprocessor_overrides
 
     if cfg.is_reward_model_training:
         preprocessor, postprocessor = make_reward_pre_post_processors(
@@ -341,7 +337,6 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
             policy_cfg=cfg.policy,
             pretrained_path=processor_pretrained_path,
             **processor_kwargs,
-            **postprocessor_kwargs,
         )
 
     if is_main_process: