PaperlayerView exact revision on arXiv

Source-linked AI summary

SCOPE: Selective Certification of Program Edits for Embodied Heuristic Learning

Junyang Shu, Bingqing Wei, Jishuo Li, Zheqi He, Yongtao Wang

arXiv:2610.23128v1cs.RO

TL;DR

Local repairs may leave structural failures unresolved, while broader edits can alter behavior that previously succeeded. SCOPE combines frontier-guided repair with replay of successful executions to assess edits. It exceeds the strongest baseline on LIBERO-PRO and Robosuite, with replay requiring a small fraction of original execution time.

  • Problem

    Local repairs may leave structural failures unresolved, while broader revisions can change decisions that supported success in other configurations, and repeated environment checks consume interaction budget.

    Source: §1 · pp. 1–2

  • Method

    SCOPE uses failed executions to guide repairs through a shared diagnostic graph and replays recorded successful traces to assess whether edits preserve behavior.

    Source: §1 · p. 2

  • Results

    80.66% on LIBERO-PRO and 84.48% on Robosuite exceed the strongest baseline by 6.51 and 3.48 percentage points, respectively.

    Source: §5 · p. 9

  • Takeaways & Limitations

    Full-trace replay takes 4.62–6.59% of corresponding original environment execution time, making historical behavior inexpensive to assess during search.

    Source: §5 · p. 9

    Repeated language-model inference remains costly, and physical-robot candidate programs may require human supervision and intervention for unsafe behavior.

    Source: §5 · p. 10

Abstract

from arXiv · show

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.

1 INTRODUCTION

SCOPE uses failed executions to guide repairs and successful executions to check whether edits preserve prior behavior. Its evaluation reports higher held-out benchmark success than the strongest baseline, alongside strong performance on real-world coordination and manipulation tasks.

1 INTRODUCTION

  • Local repairs may leave structural failures unresolved, while broader revisions can change decisions that supported success in other configurations.Repeated environment execution on historical successes can assess edits but spends interaction budget checking old behavior as well as discovering new behavior.Source: §1 · pp. 1–2
  • Failed traces guide diagnosis, while replaying successful traces checks whether edits preserve recorded API requests and terminate normally.Under faithful recording, deterministic execution, and capture of relevant external interactions, a complete match certifies success on the recorded instance; divergence identifies changed behavior for further diagnosis.Source: §1 · p. 2
  • A shared diagnostic graph marks unresolved checks with supported prerequisites as a failure frontier, guiding local repair, stage redesign, or strategy regeneration.Revisions accumulate in a shared working program across development configurations.Source: §1 · p. 2
  • After a successful configuration, SCOPE accepts the working program and, when its source changes, samples historical successes for stateful full-trace replay.For selected configurations that remain uncertified, sparse environment diagnostics provide fresh outcomes and failure evidence for the repair loop.Source: §1 · p. 2
  • 80.66% on LIBERO-PRO and 84.48% on Robosuite exceed the strongest baseline by 6.51 and 3.48 percentage points, respectively.Full-trace replay takes 4.62–6.59% of the corresponding original environment execution time.Source: §1 · p. 2
  • SCOPE shows strong real-world performance on tasks requiring base–arm coordination, bimanual coordination, and fine manipulation.The evaluation also reports improved mean held-out success over the strongest baseline on each benchmark.Source: §1 · p. 2

2 PRELIMINARY

The preliminary setup formalizes embodied heuristic learning as refining task-specific decision rules through interaction while keeping language-model parameters fixed. It also specifies the execution records and assumptions underlying replay-based certification.

2 PRELIMINARY

  • Embodied heuristic learning refines task-specific decision rules through interaction, with control programs expressing perception, conditional decisions, and actions.Given a task description and public observation and control APIs, a program maps observation histories to API calls.Source: §2 · p. 2
  • An execution instance specifies the initial state and program inputs, exogenous randomness, and API, runtime, and evaluator semantics.Execution on development reset configurations produces recorded interactions, while learning updates the program and memory without changing language-model parameters.Source: §2 · pp. 2–3
  • Replay assumes relevant external inputs and effects cross the recorded API boundary and that decisions and responses are deterministic given the execution instance.Records must faithfully preserve values, aliases, exceptions, and public-state mutations; a full match covers all external events and normal termination.Source: §2 · p. 3

3 METHOD

SCOPE combines failure-driven diagnosis with replay of selected historical successes to guide program revisions and check whether prior behavior is preserved.

3.1 FAILURE-FRONTIER MEMORY AND RETRIEVAL

  • A shared diagnostic graph represents observable checks and prerequisites; execution evidence labels checks and identifies unresolved frontier checks whose prerequisites are supported.The archive retains graph annotations and trace evidence to help locate failures and inspect their prerequisites.Source: §3.1 · p. 3

3.2 ADAPTIVE PROGRAM SEARCH

  • For each development configuration, SCOPE uses annotated executions and retrieved records to plan repairs, choosing local repair, stage redesign, or strategy regeneration.The plan states a failure hypothesis, cites evidence, identifies code or decisions to change, and describes the expected observable effect; main-search repairs are bounded per configuration.Source: §3.2 · pp. 4–5

3.3 SUCCESS-TRIGGERED UPDATES AND CERTIFICATION FEEDBACK

  • After an accepted source update, SCOPE replays selected success witnesses, restoring recorded API effects; full-tape matches with normal termination certify recorded instances under replay assumptions.A divergence stops replay, indicating changed behavior rather than task failure.Source: §3.3 · p. 5
  • Sparse diagnostics on uncertified configurations provide failure evidence for follow-up repairs, while successful trials update the accepted program and its historical witness bank before the accepted program is frozen after search.Follow-ups use the same diagnosis and update rules as main search, with a validating trial after repair.Source: §3.3 · pp. 5–6

4 EXPERIMENTS

SCOPE is evaluated on held-out simulation configurations and physical tasks, with benchmark comparisons, component ablations, and certification-rate and cost measurements. It improves on the same-backbone ASPIRE baseline in simulation and real-world evaluation.

4.1 EXPERIMENTAL SETUP

  • The evaluation covers 60 LIBERO-PRO task–setting pairs and seven Robosuite tasks, with programs frozen for held-out evaluation; physical programs are tested on 20 held-out placements per task.Simulation success is task-averaged over three independent learning and evaluation runs.Source: §4.1 · p. 6, §4.2 · p. 7

4.2 BENCHMARK RESULTS

  • 80.66 ± 1.46% overall success on LIBERO-PRO exceeds same-backbone ASPIRE by 6.51 percentage points and leads all six suite–setting combinations.The largest gains are on Goal–Task (48.80% to 61.87%) and Spatial–Pos (54.47% to 65.40%).Source: §4.2 · p. 7, Table 1 · p. 7
  • 84.48 ± 1.67% mean success on Robosuite improves over same-backbone ASPIRE by 4.29 percentage points across seven manipulation tasks.The largest gain is on two-arm lift, where success rises from 70.33% to 86.00%.Source: §4.2 · p. 7, Table 2 · p. 7
  • 66.7% overall success on the Quanta X2 robot versus ASPIRE’s 40.0% spans base–arm coordination, bimanual coordination, and fine manipulation.SCOPE improves by 30, 20, and 30 percentage points on the three tasks, respectively; cube restack remains hardest for both methods.Source: §4.2 · pp. 7–8, Table 3 · p. 8

4.3 COMPONENT ABLATIONS

  • Removing each of the three components lowers mean success in both ablation settings; failure-frontier guidance has the largest Goal–Task decrease (6.80 points), and certification feedback the largest Spatial–Pos decrease (6.07 points).The ablations cover all ten tasks in each setting with GLM-5.3-Flash.Source: §4.3 · pp. 8–9, Table 4 · p. 8

4.4 CERTIFICATION RATES AND COSTS

  • Full-trace replay certifies 38.9–60.6% of historical-success checks across LIBERO-PRO settings and 44.7% on Robosuite.The denominator retains unsupported checks; remaining checks indicate behavior changes or unavailable replay evidence.Source: §4.4 · p. 9, Figure 4 · p. 9
  • Replay takes 4.62–6.59% of original environment execution time; the full certification pipeline uses 0.54–1.36% of development time, and diagnostic executions use a further 0.87–3.10%.Sparse environment diagnostics assess the current program on selected unresolved configurations.Source: §4.4 · p. 9, Figure 4 · p. 9

5 CONCLUSION AND LIMITATIONS

SCOPE combines prerequisite-aware diagnosis with selective replay to guide program revisions and assess their effects on previously successful behavior. It outperforms the strongest baseline on both reported benchmarks, while repeated language-model inference and physical-robot safety remain limitations.

5 CONCLUSION AND LIMITATIONS

  • 80.66% on LIBERO-PRO and 84.48% on Robosuite are SCOPE’s mean held-out success rates, exceeding the strongest baseline by 6.51 and 3.48 percentage points, respectively.Full-trace replay takes 4.62–6.59% of the corresponding original environment execution time.Source: §5 · p. 9
  • SCOPE links prerequisite-aware diagnosis and selective replay of recorded successes to program revision across configurations, without updating language-model parameters.Replay checks whether accepted edits preserve recorded successful behavior; sparse environment diagnostics provide fresh outcomes for uncertified configurations.Source: §5 · p. 9, A RELATED WORK · p. 14
  • Repeated language-model inference remains costly, and early candidate programs on physical robots may require human supervision and intervention.The authors identify more selective model querying and runtime safety constraints as future directions.Source: §5 · p. 10

A RELATED WORK

  • SCOPE makes task progress explicit in a shared diagnostic graph, using prerequisite links to connect execution evidence to unresolved stages and guide repair scope.This distinguishes its approach from the embodied control and skill-learning work discussed in the related-work section.Source: A RELATED WORK · p. 14
  • SCOPE revises control programs while keeping language-model parameters fixed, and its replay can certify success on matched recorded instances under the stated assumptions.The related-work discussion contrasts this with reinforcement fine-tuning of policies.Source: A RELATED WORK · p. 14
  • SCOPE adds historical-success feedback after accepted source updates, using stateful API replay and sparse environment diagnostics to guide subsequent repairs.This is presented alongside prior feedback-driven program repair and search approaches.Source: A RELATED WORK · p. 14

B EXPERIMENTAL SETUP DETAILS

The experiments evaluate frozen programs on held-out benchmark seeds after learning and repair on development splits. The protocol also specifies implementation resources, comparison methods, and reporting across three independent runs.

B EXPERIMENTAL SETUP DETAILS

  • LIBERO-PRO covers 60 task–setting pairs across Object, Goal, and Spatial suites under position swaps and task changes, while Robosuite covers seven manipulation tasks.Programs are generated, repaired, and selected on development seeds, then evaluated on held-out seeds without further program changes.Source: B EXPERIMENTAL SETUP DETAILS · p. 15
  • Learning updates control programs and execution memory while keeping language-model parameters fixed, with GLM-5.3-Flash run locally on eight NVIDIA PRO 6000 GPUs.Perception services and simulation workloads use a separate pool of eight NVIDIA A100 GPUs.Source: B EXPERIMENTAL SETUP DETAILS · p. 15
  • Each development configuration allows at most three main-search repairs and one follow-up repair, with up to three historical success witnesses and three diagnostic trials.The witness and diagnostic limits are K = 3 and kdiag = 3, respectively.Source: B EXPERIMENTAL SETUP DETAILS · p. 15
  • Comparisons include ASPIRE with the same GLM-5.3-Flash backbone, ASPIRE’s Opus 4.6 results, and other program-based and vision-language-action baselines.Aggregate comparison scores are equal-weight means of reported suite–setting or task scores; other methods retain their original protocols.Source: B EXPERIMENTAL SETUP DETAILS · p. 15
  • The complete learning and held-out evaluation pipeline is repeated three times, and reported simulation results are means and sample standard deviations across run-level scores.Comparison methods’ original reports are point estimates.Source: B EXPERIMENTAL SETUP DETAILS · p. 15

C ABLATION PROTOCOL

The ablation protocol compares component removals under matched configurations and fixed learning limits. Each removal masks a distinct source of guidance while retaining specified parts of the repair and evaluation process.

C ABLATION PROTOCOL

  • The ablations use GLM-5.3-Flash on all ten tasks in Goal–Task and Spatial–Pos, the settings with the largest gains over the same-backbone baseline.These are the two selected settings for the Table 4 ablations.Source: C ABLATION PROTOCOL · p. 16
  • Four conditions share development and held-out configurations, with three repetitions totaling 240 task-level learning runs and the same repair, witness, and diagnostic limits.Each run independently generates its initial program and task graph and maintains its own archive and success bank.Source: C ABLATION PROTOCOL · p. 16
  • The certification-feedback ablation masks certification results, first divergences, and derived summaries from planning and code generation while retaining internal replay and diagnostics.Environment diagnostic observations and failures remain available.Source: C ABLATION PROTOCOL · p. 16
  • The failure-frontier ablation masks the diagnostic graph, node labels, failure frontiers, evidence windows, and derived summaries while retaining raw observations, traces, errors, outcomes, and source programs.Source: C ABLATION PROTOCOL · p. 16
  • The revision-level-planning ablation retains failure diagnosis and repair planning but removes the explicit choice among local repair, stage redesign, and strategy regeneration.Source: C ABLATION PROTOCOL · p. 16

D DETAILED CERTIFICATION RATES AND COSTS

Table 5 reports task macro-averaged certification rates and costs for SCOPE across LIBERO-PRO and Robosuite. Its time columns use distinct denominators, and diagnostic trial time covers environment execution only.

D DETAILED CERTIFICATION RATES AND COSTS

E REAL-WORLD TASKS AND PROTOCOL

The real-world evaluation uses a wheeled humanoid platform and three tasks spanning base–arm coordination, bimanual coordination, and fine manipulation. Each method is developed on ten configurations per task, then evaluated on 20 held-out initial placements without further repair.

E REAL-WORLD TASKS AND PROTOCOL

  • The Quanta X2 is a wheeled humanoid with two 7-DoF arms, a 4-DoF waist, a 2-DoF head, and a 2-DoF mobile base.It stands 1.64 m tall and has a nominal arm reach of 765 mm.Source: E REAL-WORLD TASKS AND PROTOCOL · p. 17
  • Each method uses ten development configurations per task, then one frozen program per task is evaluated on 20 held-out initial placements without further repair.The methods share the language model, perception modules, and robot control interfaces.Source: E REAL-WORLD TASKS AND PROTOCOL · p. 17
  • The three tasks are Cube into basket, Two arm lift, and Cube restack, covering base–arm coordination, bimanual coordination, and fine manipulation, respectively.Figure 5 shows representative execution frames progressing from left to right.Source: Figure 5 · p. 17, E REAL-WORLD TASKS AND PROTOCOL · p. 17

F RESULTS BY LIBERO TASK

This block reports individual LIBERO task results and analyzes how program edits can interfere behaviorally. It relates possible gains to execution divergence and to whether an edited conditional branch is reached.

F RESULTS BY LIBERO TASK

G BEHAVIORAL INTERFERENCE OF PROGRAM EDITS

  • Behavioral divergence probability η is the probability that two programs’ external interactions differ on the same task instance.The compared executions share initial conditions, randomness, and execution semantics.Source: G BEHAVIORAL INTERFERENCE OF PROGRAM… · p. 20
  • When edits are confined to a branch guarded by the same side-effect-free predicate, divergence is bounded by the probability that the original program enters that branch.The programs execute identically before the first branch entry, and the edited code is not executed if the branch is never entered.Source: G BEHAVIORAL INTERFERENCE OF PROGRAM… · p. 20
  • An improvement of Δ requires behavioral divergence probability of at least Δ; low divergence limits possible gains, motivating broader revisions when local repairs are insufficient.SCOPE uses execution evidence to select revision levels and replay to assess preservation on recorded instances.Source: G BEHAVIORAL INTERFERENCE OF PROGRAM… · p. 20

H.1 INSTANCE-SPECIFIC CERTIFICATION

A complete stateful replay match with normal termination certifies success on the historical witness’s recorded instance under the stated assumptions. This certification is instance-specific and does not guarantee preservation of every historical success.

H.1 INSTANCE-SPECIFIC CERTIFICATION

  • A full-tape match with normal termination certifies that the tested program succeeds on the recorded instance, under the replay assumptions.Faithful replay restores responses and associated mutations, and a full match consumes every recorded event without extra external interactions.Source: §H.1 · p. 21
  • Certification applies to the recorded instance: a fresh reset may differ, and divergent or unavailable replay does not establish task failure.Certification provides evidence for later repairs without requiring monotonic improvement or preservation of every historical success.Source: §H.1 · p. 21

H.2 FEEDBACK COST

  • Feedback-cost accounting distinguishes selected replay checks from diagnostic environment executions; the reported total also includes retrieval, verification, and self-matching.The count limits are per update, and both counts are zero when the accepted source is unchanged.Source: §H.2 · p. 21
Loading 2610.23128v1…