Auto-sync: 2026-06-28 07:42:57

dovla_cil/eval/maniskill_policy_rollout.py

@@ -63,6 +63,7 @@ def evaluate_maniskill_policy_rollout(
     retrieval_metric: str = "raw",
     retrieval_type_min_success: float = 0.0,
     retrieval_residual_scale: float = 1.0,
+    retrieval_residual_anchor: str = "expert",
     lattice_exclude_types: tuple[str, ...] = (),
 ) -> dict[str, Any]:
     """Execute a checkpoint policy from restored ManiSkill CIL states.
@@ -141,6 +142,8 @@ def evaluate_maniskill_policy_rollout(
         raise ValueError("retrieval_neighbors must be positive")
     if retrieval_metric not in {"raw", "zscore"}:
         raise ValueError("retrieval_metric must be 'raw' or 'zscore'")
+    if retrieval_residual_anchor not in {"expert", "policy"}:
+        raise ValueError("retrieval_residual_anchor must be 'expert' or 'policy'")
     if not 0.0 <= retrieval_type_min_success <= 1.0:
         raise ValueError("retrieval_type_min_success must be in [0, 1]")
     if retrieval_residual_scale < 0:
@@ -204,6 +207,8 @@ def evaluate_maniskill_policy_rollout(
                 observation_mode=model_config.observation_mode,
                 retrieval_neighbors=retrieval_neighbors,
                 retrieval_metric=retrieval_metric,
+                retrieval_type_min_success=retrieval_type_min_success,
+                retrieval_residual_anchor=retrieval_residual_anchor,
             )
     by_task: dict[str, list[_RolloutCase]] = defaultdict(list)
     for case in cases:
@@ -285,6 +290,9 @@ def evaluate_maniskill_policy_rollout(
         "retrieval_residual_scale": retrieval_residual_scale
         if selection_mode == "retrieval_residual"
         else 0.0,
+        "retrieval_residual_anchor": retrieval_residual_anchor
+        if selection_mode == "retrieval_residual"
+        else "none",
         "lattice_exclude_types": list(lattice_exclude_types),
         "policy_rollout_success_rate": _mean([row["success"] for row in rows]),
         "policy_rollout_progress": _mean([row["progress"] for row in rows]),
@@ -446,6 +454,7 @@ def _attach_retrieved_residual_candidates(
     retrieval_neighbors: int,
     retrieval_metric: str = "raw",
     retrieval_type_min_success: float = 0.0,
+    retrieval_residual_anchor: str = "expert",
 ) -> list[_RolloutCase]:
     if observation_mode != "state":
         raise ValueError("retrieval_residual currently supports state observations only")
@@ -463,7 +472,12 @@ def _attach_retrieved_residual_candidates(
         task_ids = {record.task_id for record in records}
         if len(task_ids) != 1:
             continue
-        anchor = next((record for record in records if record.candidate_type == "expert"), records[0])
+        anchor = next(
+            (record for record in records if record.candidate_type == retrieval_residual_anchor),
+            None,
+        )
+        if anchor is None:
+            anchor = next((record for record in records if record.candidate_type == "expert"), records[0])
         anchor_action = np.asarray(_numeric_action_values(anchor), dtype=np.float32)
         residuals: list[list[list[float]]] = [np.zeros_like(anchor_action).tolist()]
         candidate_types = ["policy_residual"]

logs/auto_sync_hf.log

@@ -184,3 +184,4 @@ No files have been modified since last commit. Skipping to prevent empty commit.
 No files have been modified since last commit. Skipping to prevent empty commit.
 No files have been modified since last commit. Skipping to prevent empty commit.
 No files have been modified since last commit. Skipping to prevent empty commit.
+No files have been modified since last commit. Skipping to prevent empty commit.