Embodied heuristic learning must balance strategic adaptation with the preservation of previously successful behavior. Local repairs may leave structural failures unresolved, whereas broader revisions can introduce regressions that are costly to detect through repeated environment interaction. We introduce SCOPE (Selective Certification Of Program Edits), which uses failed and successful executions as complementary evidence for program revision. Frontier-guided program search annotates a shared diagnostic graph with execution evidence to distinguish supported progress from unresolved checks. Selective behavioral certification uses stateful API replay to restore recorded responses and mutations, checking whether edits preserve selected historical successes. Diagnosis and certification feedback, supplemented by sparse environment diagnostics, guide consecutive repairs across configurations. SCOPE outperforms the strongest baseline in both simulation and real-world tasks. On LIBERO-PRO and Robosuite, it achieves mean success rates of 80.66% and 84.48%, with respective gains of 6.51 and 3.48 percentage points. Our code is available at https://github.com/VDIGPKU/SCOPE.
Figures & tables
Figure 1: Overview of SCOPE. Execution evidence annotates a shared diagnostic graph to guide local repair, stage redesign, or strategy regeneration. Success updates the accepted program and historical success bank. When the accepted source changes, selected historical tapes are replayed by a local executor. Full matches certify the recorded instances under the replay assumptions. Sparse environment diagnostics on uncertified configurations supply failure evidence for subsequent repair. The accepted program is frozen after search.
Figure 2: Stateful full-trace replay. (a) A local executor compares each API request and pre-call state with the tape, reproduces the recorded response and mutations while preserving shared references, and resumes the program. Omitting a mutation changes the next request in this illustrative example. (b) Matching every event, making no extra external calls, and terminating normally certifies success on the recorded instance ξw under the replay assumptions. A changed argument stops replay at the first divergence, leaving the instance uncertified.
Method
Object
Goal
Spatial
Overall
Pos
Task
Pos
Task
Pos
Task
Pos
Task
All
OpenVLA
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
π0
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
SUREFlow
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
MolmoAct
6.00
0.00
0.00
0.00
0.00
0.00
2.00
0.00
1.00
X-VLA
2.00
8.00
1.00
9.00
0.00
0.00
1.00
5.67
3.33
Table 1: LIBERO-PRO results. Success rates (%) under position swaps (Pos) and task changes (Task). Overall averages over suites, and All averages over both settings.
Method
Cube lift
Cube stack
Cube restack
Spill wipe
Two arm handover
Two arm lift
Nut assembly
Mean
CaP-Agent0
97.00
98.00
89.00
100.00
20.00
74.00
0.00
68.29
RHO †
98.00
98.00
54.00
100.00
86.00
53.00
1.00
70.00
ASPIRE Opus 4.6
97.00
99.00
100.00
99.00
92.00
71.00
9.00
81.00
ASPIRE GLM-5.3-Flash
98.33±1.53
97.33±1.53
97.67±1.53
98.33±1.15
88.67±5.13
70.33±6.81
10.67±4.73
80.19±2.14
SCOPE GLM-5.3-Flash
100.00±0.00
99.00±1.00
100.00±0.00
100.00±0.00
91.67±3.21
86.00±3.61
14.67±4.93
84.48±1.67
Table 2: Robosuite results. Success rates (%) on seven manipulation tasks. Mean averages over tasks.
Task
Capability
ASPIRE
SCOPE
Cube into basket
Base–arm coordination
9/20 (45.0%)
15/20 (75.0%)
Two arm lift
Bimanual coordination
10/20 (50.0%)
14/20 (70.0%)
Cube restack
Fine manipulation
5/20 (25.0%)
11/20 (55.0%)
Overall
—
24/60 (40.0%)
40/60 (66.7%)
Table 3: Real-world generalization. Successes/trials (success rate) for frozen programs on 20 held-out physical configurations per task. Overall pools 60 trials per method. Bold marks the higher success rate.
Figure 3: Real-world repair of a cube-in-basket task. (a) Gripper occlusion prompts backward motion for localization. (b) The target is visible but out of reach, prompting forward motion. (c) Renewed occlusion motivates caching the target position before approaching. (d) The revised program completes the task. Each panel pairs a key frame with the corresponding repair or outcome.
Method / variant
Goal–Task
Spatial–Pos
ASPIRE (baseline)
48.80±11.68
54.47±3.45
Full SCOPE
61.87±5.41
65.40±3.82
w/o certification feedback
57.60±7.33
59.33±4.16
w/o failure-frontier guidance
55.07±9.10
59.93±3.72
w/o explicit revision-level planning
58.87±7.11
62.13±3.88
Table 4: Component ablations on LIBERO-PRO. Success rates (%) on all 10 tasks in each setting with GLM-5.3-Flash, reported as mean ± sample standard deviation over three runs.
Figure 4: Certification rates and feedback costs of SCOPE. Each point is a task macro-average for one setting with GLM-5.3-Flash, covering 60 LIBERO-PRO task–setting pairs and seven Robosuite tasks. (a) Certification rate versus replay time relative to the corresponding original environment execution. (b) Certification-pipeline and diagnostic environment-execution times, each relative to total development learning time. Labels give (x,y) in percent. Circles and triangles denote LIBERO-PRO Pos and Task settings, while stars denote Robosuite. The certification pipeline includes replay, retrieval, verification, and self-matching. Diagnostic time covers environment execution only.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Setting
Certification rate
Replay time / original execution time
Certification time / total learning time
Diagnostic trial time / total learning time
Object–Pos
60.6
4.93
0.54
0.87
Object–Task
55.3
5.53
0.83
1.41
Goal–Pos
49.0
6.12
0.75
1.28
Goal–Task
38.9
6.59
1.36
2.67
Spatial–Pos
42.1
6.25
1.04
2.04
Spatial–Task
41.9
5.94
0.98
2.12
Appendix
Table 5: Certification rates and costs of SCOPE. Task macro-averages (%) with GLM-5.3-Flash over 60 LIBERO-PRO task-setting pairs and seven Robosuite tasks.
Figure 5: Real-world task executions. Representative frames progress from left to right for Cube into basket (top), Two arm lift (middle), and Cube restack (bottom). The tasks evaluate mobile manipulation with base–arm coordination, bimanual coordination, and fine manipulation, respectively.
Task
ASPIRE Opus 4.6
ASPIRE GLM-5.3-Flash
SCOPE GLM-5.3-Flash
LIBERO-Object
Alphabet soup
100.00
98.67±1.15
100.00±0.00
BBQ sauce
100.00
99.33±1.15
100.00±0.00
Butter
100.00
98.67±2.31
99.33±1.15
Chocolate pudding
100.00
98.67±2.31
100.00±0.00
Cream cheese
100.00
98.00±2.00
100.00±0.00
Appendix
Table 6: Success rates (%) for individual LIBERO tasks under the Pos setting.
Task
ASPIRE Opus 4.6
ASPIRE GLM-5.3-Flash
SCOPE GLM-5.3-Flash
LIBERO-Object
Alphabet soup
100.00
98.00±2.00
100.00±0.00
BBQ sauce
100.00
99.33±1.15
100.00±0.00
Butter
98.00
98.67±2.31
100.00±0.00
Chocolate pudding
100.00
98.67±1.15
99.33±1.15
Cream cheese
72.00
84.00±5.29
91.33±4.16
Appendix
Table 7: Success rates (%) for individual LIBERO tasks under the Task setting.
An agent can retain successful trajectories without learning which part of an experience should repair a new discrepancy. We formulate passive imitation (PI) as an architectural loop that localizes a violated contract, extracts a local goal, proposes an effect-matched skill, tests a small patch, and retains evidence after full-task verification. We instantiate this loop as a deterministic symbolic controller and analyze its finite repair-query complexity under explicit identifiability and noninterference assumptions. Across ten simulated tool worlds, PI solves all 600 cold held-out composition tasks with three allowed small trials, compared with 6.17% for a whole-trajectory control, while an equally expressive compositional repair control matches PI. On an unseen tool revision, PI achieves 70.0% success versus 24.0% without small trials and 79.0% with self-trial enumeration; an exploratory 24-task adaptation stream subsequently raises first-submission success to 100%, supporting a limited correction-and-retention mechanism rather than a unique incentive advantage or general intelligence. The CPU-only implementation, integrity checks, and reproducible results are available at https://github.com/FriendOfYishan/PI.
Kai Hu, Hua Huang
Beijing International Studies University · Chongqing University of Posts and Telecommunications
An embodied brain, or System 2, must turn an abstract goal into actions while discovering relevant scene conditions and checking whether executed skills achieved their intended effects. Open-loop evaluations cannot test observations shaped by an agent's decisions, while some closed-loop evaluations provide limited evidence during execution or assume idealized skill outcomes. We introduce SCOPER, a suite for studying observation-grounded planning, execution, and recovery. SCOPER-bench combines high-level goals, ongoing visual observations, a shared rule-based System 1 for low-level actions, and controlled execution failures. Its construction and validation pipeline yields 870 tasks across BEHAVIOR and ProcTHOR, complemented by about 3,500 situated questions tied to the evaluated agent's execution. Benchmark results reveal difficulties in retaining scene evidence, monitoring progress, and recovering from unexpected outcomes. To address these difficulties, SCOPER-harness combines Evidence Memory and Progress Verification with a dynamic task agenda and bounded exploration, without additional model training. Across two evaluated open-source backbones and two simulators, the harness improves task success by 11.1-28.6%; a harnessed 9B model outperforms a directly executed 27B model on both simulators. In an offline next-action evaluation using observations from 200 real-robot demonstrations, the harness also improves full-call accuracy by 14.7% and 15.3%. SCOPER provides a setting for diagnosing and improving embodied reasoning throughout a complete task execution cycle.
Yuxi Wei, Ye Fang, Shuai Yang +3
The University of Hong Kong · Morphi Robot · Fudan University +1
Generate--test--revise loops are common in coding agents, but repetition alone provides no reliability guarantee. We study the gap between finding a correct patch and retaining, verifying, and submitting it. A sealed five-seed study over 30 HumanEval repairs produces 900 three-revision trajectories. Under forced revision, current correctness with current traces falls from 0.820 after one revision to 0.673 after two, although ever-correct rises to 0.847. Two common-state studies use 2,430 branches from identical frozen programs to remove post-treatment risk-set bias. In a prespecified 14B replication, stale traces harm 34/135 correct starts versus 4/135 with current traces, a 22.2-point increase (task-cluster 95% CI [8.9,37.0], exact Holm p=0.0337). A prospective 540-rollout policy eliminates observed correct-start harm but reduces wrong-start repair and fails its joint criterion. Repository experiments over 24 bugs and four coder stacks expose floor effects and component heterogeneity without Holm-significant effects. We therefore separate admission, preservation, grounded certification, competence, and liveness. We derive an evidence-bound typed loop contract and instantiate its mechanically enforceable subset in a reference implementation that binds verifier evidence to exact code states, preserves verified checkpoints, and emits auditable admission receipts. The implementation is an executable specification and conformance artifact, not evidence of improved repair competence or calibrated verifier dependence.