auto-sync 2026-07-02T15:44:03Z workspace (part 4)

workspace/tests/test_maniskill_policy_rollout.py

@@ -1236,6 +1236,32 @@ def test_retrieval_residual_candidate_type_bonus_repeats_over_scale_grid() -> No
     assert index.tolist() == [1]
 
 
+def test_retrieval_residual_scale_bonus_selects_scaled_tangent() -> None:
+    import torch
+
+    mean = torch.zeros(1, 1, 1)
+    residual = torch.ones_like(mean)
+    model = _StubModel(torch, mean, best_offset=torch.zeros_like(mean))
+    residuals = residual.unsqueeze(1)
+    actions, index = _select_action_chunk(
+        model,
+        observations=torch.zeros(1, 3),
+        instructions=["a"],
+        torch=torch,
+        selection_mode="retrieval_residual",
+        num_candidates=1,
+        candidate_sigma=0.0,
+        selection_seed=0,
+        action_candidates=residuals,
+        retrieval_residual_scale=1.0,
+        retrieval_residual_scales=(0.25, 0.5),
+        residual_scale_bonuses={"0.5": 2.0},
+    )
+
+    assert torch.allclose(actions, mean + 0.5 * residual)
+    assert index.tolist() == [1]
+
+
 def test_retrieval_residual_action_l2_penalty_prefers_shorter_tangent() -> None:
     import torch
 

workspace/tests/test_oracle_selector_calibration.py

@@ -17,6 +17,7 @@ def test_build_oracle_selector_calibration_from_branch_scores(tmp_path) -> None:
                         "candidate_oracle_branch_progress": [0.2, 0.8],
                         "candidate_oracle_branch_successes": [False, False],
                         "candidate_oracle_types": ["policy", "residual_good"],
+                        "candidate_oracle_residual_scales": [0.25, 0.5],
                         "candidate_oracle_valid_mask": [True, True],
                     },
                     {
@@ -25,6 +26,7 @@ def test_build_oracle_selector_calibration_from_branch_scores(tmp_path) -> None:
                         "candidate_oracle_branch_progress": [0.4, 0.6],
                         "candidate_oracle_branch_successes": [False, False],
                         "candidate_oracle_types": ["policy", "residual_good"],
+                        "candidate_oracle_residual_scales": [0.25, 0.5],
                         "candidate_oracle_valid_mask": [True, True],
                     },
                     {
@@ -33,6 +35,7 @@ def test_build_oracle_selector_calibration_from_branch_scores(tmp_path) -> None:
                         "candidate_oracle_branch_progress": [0.9, 0.1],
                         "candidate_oracle_branch_successes": [False, False],
                         "candidate_oracle_types": ["policy", "residual_bad"],
+                        "candidate_oracle_residual_scales": [0.25, 0.5],
                         "candidate_oracle_valid_mask": [True, True],
                     },
                 ]
@@ -46,15 +49,19 @@ def test_build_oracle_selector_calibration_from_branch_scores(tmp_path) -> None:
         max_rank=2,
         rank_scale=0.1,
         type_scale=0.1,
+        scale_scale=0.1,
         min_count=1,
         max_abs_rank_bias=None,
         max_abs_type_bonus=None,
+        max_abs_scale_bonus=None,
     )
 
     pick_rank_biases = result["field_rank_biases_by_task"]["PickCube-v1"]
     stack_rank_biases = result["field_rank_biases_by_task"]["StackCube-v1"]
     pick_type_bonuses = result["candidate_type_bonuses_by_task"]["PickCube-v1"]
     stack_type_bonuses = result["candidate_type_bonuses_by_task"]["StackCube-v1"]
+    pick_scale_bonuses = result["residual_scale_bonuses_by_task"]["PickCube-v1"]
+    stack_scale_bonuses = result["residual_scale_bonuses_by_task"]["StackCube-v1"]
 
     assert pick_rank_biases[0] == 0.0
     assert abs(pick_rank_biases[1] - 0.04) < 1e-12
@@ -62,3 +69,5 @@ def test_build_oracle_selector_calibration_from_branch_scores(tmp_path) -> None:
     assert abs(stack_rank_biases[1] + 0.08) < 1e-12
     assert pick_type_bonuses["residual_good"] > pick_type_bonuses["policy"]
     assert stack_type_bonuses["residual_bad"] < stack_type_bonuses["policy"]
+    assert pick_scale_bonuses["0.5"] > pick_scale_bonuses["0.25"]
+    assert stack_scale_bonuses["0.5"] < stack_scale_bonuses["0.25"]