Auto-sync: 2026-06-29 01:34:49 (part 3)

scripts/build_paper_table_status.py

@@ -505,6 +505,36 @@ SPECS = [
         story_role="strong train-source progress prior replacing typed no-op prior",
         pending_job="14894675/14894677",
     ),
+    ResultSpec(
+        key="retrieval_residual_k4_mean_srcscorebonus0015",
+        label="K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, source-score bonus 0.015",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_srcscorebonus0p015_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="train-source reward-score prior for sparse residual transport",
+        pending_job="14897123/14897126",
+    ),
+    ResultSpec(
+        key="retrieval_residual_k4_mean_srcscorebonus002",
+        label="K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, source-score bonus 0.02",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_srcscorebonus0p02_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="train-source reward-score prior for sparse residual transport",
+        pending_job="14897124/14897127",
+    ),
+    ResultSpec(
+        key="retrieval_residual_k4_mean_srcscorebonus0025",
+        label="K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, source-score bonus 0.025",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_srcscorebonus0p025_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="train-source reward-score prior for sparse residual transport",
+        pending_job="14897125/14897128",
+    ),
     ResultSpec(
         key="retrieval_residual_taskrelative_k4_mean_noopbonus003",
         label="K4 task-relative mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.03",

scripts/eval_maniskill_policy_rollout.py

@@ -143,6 +143,13 @@ def main(argv: list[str] | None = None) -> int:
         help="Scale for adding a train-source progress prior to each retrieved residual "
         "candidate before field selection. The policy_residual fallback receives zero.",
     )
+    parser.add_argument(
+        "--retrieval-residual-source-score-bonus-scale",
+        type=float,
+        default=0.0,
+        help="Scale for adding a train-source reward-score prior to each retrieved residual "
+        "candidate before field selection. Score is progress plus terminal success.",
+    )
     parser.add_argument(
         "--retrieval-residual-scale",
         type=float,
@@ -238,6 +245,9 @@ def main(argv: list[str] | None = None) -> int:
         retrieval_residual_source_progress_bonus_scale=(
             args.retrieval_residual_source_progress_bonus_scale
         ),
+        retrieval_residual_source_score_bonus_scale=(
+            args.retrieval_residual_source_score_bonus_scale
+        ),
         retrieval_residual_scale=args.retrieval_residual_scale,
         retrieval_residual_scales=retrieval_residual_scales,
         retrieval_residual_anchor=args.retrieval_residual_anchor,

scripts/slurm/eval_maniskill_policy_rollout.sbatch

@@ -55,6 +55,7 @@ RETRIEVAL_METRIC="${RETRIEVAL_METRIC:-raw}"
 RETRIEVAL_TYPE_MIN_SUCCESS="${RETRIEVAL_TYPE_MIN_SUCCESS:-0.0}"
 RETRIEVAL_RESIDUAL_MIN_SOURCE_PROGRESS="${RETRIEVAL_RESIDUAL_MIN_SOURCE_PROGRESS:-0.0}"
 RETRIEVAL_RESIDUAL_SOURCE_PROGRESS_BONUS_SCALE="${RETRIEVAL_RESIDUAL_SOURCE_PROGRESS_BONUS_SCALE:-0.0}"
+RETRIEVAL_RESIDUAL_SOURCE_SCORE_BONUS_SCALE="${RETRIEVAL_RESIDUAL_SOURCE_SCORE_BONUS_SCALE:-0.0}"
 RETRIEVAL_RESIDUAL_SCALE="${RETRIEVAL_RESIDUAL_SCALE:-1.0}"
 RETRIEVAL_RESIDUAL_SCALES="${RETRIEVAL_RESIDUAL_SCALES:-}"
 if [[ -n "${RETRIEVAL_RESIDUAL_SCALES_COLON:-}" ]]; then
@@ -122,6 +123,7 @@ apptainer exec --nv \
     --retrieval-type-min-success "$RETRIEVAL_TYPE_MIN_SUCCESS" \
     --retrieval-residual-min-source-progress "$RETRIEVAL_RESIDUAL_MIN_SOURCE_PROGRESS" \
     --retrieval-residual-source-progress-bonus-scale "$RETRIEVAL_RESIDUAL_SOURCE_PROGRESS_BONUS_SCALE" \
+    --retrieval-residual-source-score-bonus-scale "$RETRIEVAL_RESIDUAL_SOURCE_SCORE_BONUS_SCALE" \
     --retrieval-residual-scale "$RETRIEVAL_RESIDUAL_SCALE" \
     --retrieval-residual-scales "$RETRIEVAL_RESIDUAL_SCALES" \
     --retrieval-residual-anchor "$RETRIEVAL_RESIDUAL_ANCHOR" \

scripts/slurm/eval_maniskill_policy_rollout_cpu_smoke.sbatch

@@ -54,6 +54,7 @@ RETRIEVAL_METRIC="${RETRIEVAL_METRIC:-raw}"
 RETRIEVAL_TYPE_MIN_SUCCESS="${RETRIEVAL_TYPE_MIN_SUCCESS:-0.0}"
 RETRIEVAL_RESIDUAL_MIN_SOURCE_PROGRESS="${RETRIEVAL_RESIDUAL_MIN_SOURCE_PROGRESS:-0.0}"
 RETRIEVAL_RESIDUAL_SOURCE_PROGRESS_BONUS_SCALE="${RETRIEVAL_RESIDUAL_SOURCE_PROGRESS_BONUS_SCALE:-0.0}"
+RETRIEVAL_RESIDUAL_SOURCE_SCORE_BONUS_SCALE="${RETRIEVAL_RESIDUAL_SOURCE_SCORE_BONUS_SCALE:-0.0}"
 RETRIEVAL_RESIDUAL_SCALE="${RETRIEVAL_RESIDUAL_SCALE:-1.0}"
 RETRIEVAL_RESIDUAL_SCALES="${RETRIEVAL_RESIDUAL_SCALES:-}"
 if [[ -n "${RETRIEVAL_RESIDUAL_SCALES_COLON:-}" ]]; then
@@ -118,6 +119,7 @@ apptainer exec \
     --retrieval-type-min-success "$RETRIEVAL_TYPE_MIN_SUCCESS" \
     --retrieval-residual-min-source-progress "$RETRIEVAL_RESIDUAL_MIN_SOURCE_PROGRESS" \
     --retrieval-residual-source-progress-bonus-scale "$RETRIEVAL_RESIDUAL_SOURCE_PROGRESS_BONUS_SCALE" \
+    --retrieval-residual-source-score-bonus-scale "$RETRIEVAL_RESIDUAL_SOURCE_SCORE_BONUS_SCALE" \
     --retrieval-residual-scale "$RETRIEVAL_RESIDUAL_SCALE" \
     --retrieval-residual-scales "$RETRIEVAL_RESIDUAL_SCALES" \
     --retrieval-residual-anchor "$RETRIEVAL_RESIDUAL_ANCHOR" \

scripts/slurm/smoke_retrieval_metric_unit.sbatch

@@ -172,14 +172,23 @@ case_task_relative = _RolloutCase(
 )
 assert task_relative_attached.candidate_source_group_id == "train_actor_match", task_relative_attached
 
-def record_progress(group_id, candidate_type, action_value, progress, feature):
+def record_progress(
+    group_id,
+    candidate_type,
+    action_value,
+    progress,
+    feature,
+    terminal_success=None,
+):
+    if terminal_success is None:
+        terminal_success = progress >= 1.0
     return SimpleNamespace(
         group_id=group_id,
         task_id="PickCube-v1",
         candidate_type=candidate_type,
         record_id=f"{group_id}-{candidate_type}-{action_value}",
         observation_inline={"features": feature},
-        reward=SimpleNamespace(progress=progress, terminal_success=progress >= 1.0),
+        reward=SimpleNamespace(progress=progress, terminal_success=terminal_success),
         action_chunk=ActionChunk(
             representation="continuous",
             horizon=1,
@@ -190,13 +199,27 @@ def record_progress(group_id, candidate_type, action_value, progress, feature):
 groups_progress = {
     "train_a": [
         record_progress("train_a", "expert", 1.0, 1.0, [0.0, 0.0]),
-        record_progress("train_a", "no_op", 1.2, 0.8, [0.0, 0.0]),
+        record_progress(
+            "train_a",
+            "no_op",
+            1.2,
+            0.8,
+            [0.0, 0.0],
+            terminal_success=True,
+        ),
         record_progress("train_a", "wrong_gripper", 1.4, 0.2, [0.0, 0.0]),
     ],
     "train_b": [
         record_progress("train_b", "expert", 2.0, 1.0, [10.0, 0.0]),
         record_progress("train_b", "no_op", 2.2, 0.8, [10.0, 0.0]),
-        record_progress("train_b", "wrong_gripper", 2.4, 0.8, [10.0, 0.0]),
+        record_progress(
+            "train_b",
+            "wrong_gripper",
+            2.4,
+            0.8,
+            [10.0, 0.0],
+            terminal_success=True,
+        ),
     ],
     "heldout_a": [record_progress("heldout_a", "expert", 9.0, 1.0, [0.0, 0.0])],
     "heldout_b": [record_progress("heldout_b", "expert", 9.0, 1.0, [10.0, 0.0])],
@@ -232,6 +255,7 @@ progress_attached = _attach_retrieved_residual_candidates(
     retrieval_neighbors=1,
     retrieval_residual_min_source_progress=0.5,
     retrieval_residual_source_progress_bonus_scale=0.1,
+    retrieval_residual_source_score_bonus_scale=0.05,
 )
 assert [len(case.candidate_action_values) for case in progress_attached] == [3, 3], progress_attached
 assert progress_attached[0].candidate_types == [
@@ -244,8 +268,8 @@ assert progress_attached[1].candidate_types == [
     "residual_no_op",
     "residual_wrong_gripper",
 ], progress_attached[1].candidate_types
-assert np.allclose(progress_attached[0].candidate_score_bonuses, [0.0, 0.08, 0.0])
-assert np.allclose(progress_attached[1].candidate_score_bonuses, [0.0, 0.08, 0.08])
+assert np.allclose(progress_attached[0].candidate_score_bonuses, [0.0, 0.17, 0.0])
+assert np.allclose(progress_attached[1].candidate_score_bonuses, [0.0, 0.12, 0.17])
 print({
     "status": "ok",
     "raw": raw_attached.candidate_source_group_id,

scripts/slurm/summarize_h16_policy_ckpt.sbatch

@@ -76,6 +76,9 @@ for result_path in sorted(base_dir.glob(f"seed_*/{out_name}")):
             "retrieval_residual_source_progress_bonus_scale": data.get(
                 "retrieval_residual_source_progress_bonus_scale", 0.0
             ),
+            "retrieval_residual_source_score_bonus_scale": data.get(
+                "retrieval_residual_source_score_bonus_scale", 0.0
+            ),
             "retrieval_residual_scale": data.get("retrieval_residual_scale", 0.0),
             "retrieval_residual_scales": data.get("retrieval_residual_scales", []),
             "retrieval_residual_anchor": data.get("retrieval_residual_anchor", "none"),
@@ -142,14 +145,14 @@ lines = [
     f"Mean progress: {summary['mean_progress']:.2%}",
     f"Mean action MSE to best: {summary['mean_action_mse_to_best']:.3f}",
     "",
-        "| seed | mode | k | policy cand | retrieval K | retrieval metric | residual anchor | residual reduce | min type success | min source progress | source progress bonus | residual scale | residual scales | margin | sigma | opt steps | trust | success | progress | oracle | action MSE |",
-        "|---:|---|---:|---|---:|---|---|---|---:|---:|---:|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|",
+        "| seed | mode | k | policy cand | retrieval K | retrieval metric | residual anchor | residual reduce | min type success | min source progress | source progress bonus | source score bonus | residual scale | residual scales | margin | sigma | opt steps | trust | success | progress | oracle | action MSE |",
+        "|---:|---|---:|---|---:|---|---|---|---:|---:|---:|---:|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|",
 ]
 for row in rows:
     scales = row.get("retrieval_residual_scales") or []
     scale_grid = ",".join(f"{float(scale):.2f}" for scale in scales) if scales else "none"
     lines.append(
-        "| {seed} | {mode} | {k} | {policy_cand} | {retrieval} | {metric} | {anchor} | {reduce} | {min_success:.2f} | {min_source_progress:.2f} | {source_progress_bonus:.3f} | {scale:.2f} | {scale_grid} | {margin:.3f} | {sigma:.2f} | {steps} | {trust:.2f} | "
+        "| {seed} | {mode} | {k} | {policy_cand} | {retrieval} | {metric} | {anchor} | {reduce} | {min_success:.2f} | {min_source_progress:.2f} | {source_progress_bonus:.3f} | {source_score_bonus:.3f} | {scale:.2f} | {scale_grid} | {margin:.3f} | {sigma:.2f} | {steps} | {trust:.2f} | "
         "{success:.2%} | {progress:.2%} | {oracle:.2%} | {mse:.3f} |".format(
             seed=row["seed"],
             mode=row.get("selection_mode") or "policy",
@@ -165,6 +168,10 @@ for row in rows:
                 "retrieval_residual_source_progress_bonus_scale"
             )
             or 0.0,
+            source_score_bonus=row.get(
+                "retrieval_residual_source_score_bonus_scale"
+            )
+            or 0.0,
             scale=row.get("retrieval_residual_scale") or 0.0,
             scale_grid=scale_grid,
             margin=row.get("selection_margin") or 0.0,

tests/test_maniskill_policy_rollout.py

@@ -1093,14 +1093,17 @@ def test_retrieval_residual_source_progress_threshold_filters_individual_residua
         candidate_type: str,
         action_value: float,
         progress: float,
+        terminal_success: bool | None = None,
     ):
+        if terminal_success is None:
+            terminal_success = progress >= 1.0
         return SimpleNamespace(
             group_id=group_id,
             task_id="PickCube-v1",
             candidate_type=candidate_type,
             record_id=f"{group_id}-{candidate_type}-{action_value}",
             observation_inline={"features": [0.0, 0.0]},
-            reward=SimpleNamespace(progress=progress, terminal_success=progress >= 1.0),
+            reward=SimpleNamespace(progress=progress, terminal_success=terminal_success),
             action_chunk=ActionChunk(
                 representation="continuous",
                 horizon=1,
@@ -1111,7 +1114,7 @@ def test_retrieval_residual_source_progress_threshold_filters_individual_residua
     groups = {
         "train_a": [
             record("train_a", "expert", 1.0, 1.0),
-            record("train_a", "no_op", 1.2, 0.8),
+            record("train_a", "no_op", 1.2, 0.8, terminal_success=True),
             record("train_a", "wrong_gripper", 1.4, 0.2),
         ],
         "heldout": [
@@ -1148,10 +1151,11 @@ def test_retrieval_residual_source_progress_threshold_filters_individual_residua
         retrieval_neighbors=1,
         retrieval_residual_min_source_progress=0.5,
         retrieval_residual_source_progress_bonus_scale=0.1,
+        retrieval_residual_source_score_bonus_scale=0.05,
     )
 
     assert attached.candidate_types == ["policy_residual", "residual_no_op"]
-    assert np.allclose(attached.candidate_score_bonuses, [0.0, 0.08])
+    assert np.allclose(attached.candidate_score_bonuses, [0.0, 0.17])
     assert np.allclose(
         np.asarray(attached.candidate_action_values, dtype=np.float32),
         np.asarray([[[0.0, 0.0]], [[0.2, 0.0]]]),
@@ -1165,14 +1169,17 @@ def test_retrieval_residual_source_progress_padding_keeps_batches_rectangular()
         action_value: float,
         progress: float,
         feature: list[float],
+        terminal_success: bool | None = None,
     ):
+        if terminal_success is None:
+            terminal_success = progress >= 1.0
         return SimpleNamespace(
             group_id=group_id,
             task_id="PickCube-v1",
             candidate_type=candidate_type,
             record_id=f"{group_id}-{candidate_type}-{action_value}",
             observation_inline={"features": feature},
-            reward=SimpleNamespace(progress=progress, terminal_success=progress >= 1.0),
+            reward=SimpleNamespace(progress=progress, terminal_success=terminal_success),
             action_chunk=ActionChunk(
                 representation="continuous",
                 horizon=1,
@@ -1183,13 +1190,20 @@ def test_retrieval_residual_source_progress_padding_keeps_batches_rectangular()
     groups = {
         "train_a": [
             record("train_a", "expert", 1.0, 1.0, [0.0, 0.0]),
-            record("train_a", "no_op", 1.2, 0.8, [0.0, 0.0]),
+            record("train_a", "no_op", 1.2, 0.8, [0.0, 0.0], terminal_success=True),
             record("train_a", "wrong_gripper", 1.4, 0.2, [0.0, 0.0]),
         ],
         "train_b": [
             record("train_b", "expert", 2.0, 1.0, [10.0, 0.0]),
             record("train_b", "no_op", 2.2, 0.8, [10.0, 0.0]),
-            record("train_b", "wrong_gripper", 2.4, 0.8, [10.0, 0.0]),
+            record(
+                "train_b",
+                "wrong_gripper",
+                2.4,
+                0.8,
+                [10.0, 0.0],
+                terminal_success=True,
+            ),
         ],
         "heldout_a": [record("heldout_a", "expert", 9.0, 1.0, [0.0, 0.0])],
         "heldout_b": [record("heldout_b", "expert", 9.0, 1.0, [10.0, 0.0])],
@@ -1227,6 +1241,7 @@ def test_retrieval_residual_source_progress_padding_keeps_batches_rectangular()
         retrieval_neighbors=1,
         retrieval_residual_min_source_progress=0.5,
         retrieval_residual_source_progress_bonus_scale=0.1,
+        retrieval_residual_source_score_bonus_scale=0.05,
     )
 
     assert [len(case.candidate_action_values) for case in attached] == [3, 3]
@@ -1240,5 +1255,5 @@ def test_retrieval_residual_source_progress_padding_keeps_batches_rectangular()
         "residual_no_op",
         "residual_wrong_gripper",
     ]
-    assert np.allclose(attached[0].candidate_score_bonuses, [0.0, 0.08, 0.0])
-    assert np.allclose(attached[1].candidate_score_bonuses, [0.0, 0.08, 0.08])
+    assert np.allclose(attached[0].candidate_score_bonuses, [0.0, 0.17, 0.0])
+    assert np.allclose(attached[1].candidate_score_bonuses, [0.0, 0.12, 0.17])