PaperlayerView exact revision on arXiv

Source-linked AI summary

RoboHarn-Evo: Evolving Hierarchical Physical Knowledge for Self-Improving Robotic Manipulation

Shifeng Bao, Fanding Huang, Yihan Lin, Youhe Feng, Guanlin Li, Chen Zhao, Yang Li, Jiawei He, Cheng Chi, Jing Zhang

arXiv:2609.37583v2cs.RO

TL;DR

Reliable improvement from physical experience without retraining the base model remains a challenge for manipulation. RoboHarn-Evo evolves hierarchical task and action knowledge from interaction; on RMBench, HPK improves average success by up to 24.2 percentage points across agent models.

  • Problem

    A central challenge is turning accumulated physical interaction into reliable improvement in subsequent behavior without repeatedly retraining the underlying model.

    Source: §1 · p. 2

  • Method

    RoboHarn-Evo uses physical feedback to evolve Task Knowledge about subtask choice and completion and Action Knowledge about object-relative strategies and physical effects, with the base VLM and executor fixed.

    Source: §1 · p. 2

  • Results

    On RMBench, HPK improves average success by up to 24.2 percentage points across different agent models.

    Source: §1 · p. 3

  • Takeaways & Limitations

    Frozen HPK transferred zero-shot from RMBench to RoboDojo, improving average success by 35.0 and 25.0 percentage points with GPT-5.5 and GPT-6, respectively.

    Source: §1 · p. 3

Abstract

from arXiv · show

Vision-language models can coordinate long-horizon robot manipulation, yet successful task reasoning still depends on whether local physical interactions produce the intended effects. We study how repeated interaction can improve this capability without updating the base model. We introduce RoboHarn-Evo, a dual-loop harness that evolves Hierarchical Physical Knowledge (HPK) from physical experience. HPK couples two levels of reusable knowledge: Task Knowledge captures which subtask should be executed and when it is complete, while Action Knowledge captures object-relative geometric strategies and their physical effects. During execution, the agent retrieves knowledge at the corresponding decision level and grounds it in the current scene under the task goal. Across episodes, physical feedback is used to revise historical knowledge, update its applicability, and organize reusable entries for subsequent retrieval. Experiments on RMBench show that HPK improves average success by up to 24.2 percentage points across different agent models. With 80 interaction rollouts, held-out success rises from 48.3% to 75.0% for GPT-5.5 and from 70.0% to 88.3% for GPT-6. RoboHarn-Evo also resolves over 83% of historical knowledge errors while retaining 95.8% of valid knowledge, and transfers zero-shot from RMBench to RoboDojo with gains of 35.0 and 25.0 percentage points. These results demonstrate that physical interaction can be accumulated into reusable knowledge for improving subsequent manipulation.

1 INTRODUCTION

The section frames reliable improvement from accumulated physical experience as a central challenge, especially because task completion and local action effects require different checks. RoboHarn-Evo evolves both kinds of knowledge while holding the base model and executor fixed, and evaluates knowledge correction and held-out behavioral improvement.

1 INTRODUCTION

  • Task-level completion and action-level effects need separate criteria: a successful grasp does not establish that a cover was placed at its intended location.The examples include switching blocks before confirming completion and releasing a cover off target.Source: §1 · p. 2
  • RoboHarn-Evo evolves Task Knowledge for subtask choice and completion and Action Knowledge for object-relative strategies and physical effects, checking each at its corresponding level.It uses recorded observations, tool calls, and outcomes to revise knowledge and determine its retrieval eligibility while keeping the base VLM and executor fixed.Source: §1 · p. 2
  • The evaluation tests whether physical evidence corrects or deactivates erroneous entries while preserving valid ones, and whether frozen knowledge improves held-out performance without evaluation feedback entering updates.Cross-domain transfer is also evaluated.Source: §1 · p. 2
  • After 80 interaction rollouts, held-out success rises from 48.3% to 75.0% for GPT-5.5 and from 70.0% to 88.3% for GPT-6; RMBench average success improves by up to 24.2 percentage points.The evaluation also reports zero-shot transfer gains on RoboDojo.Source: §1 · p. 3
  • Across RMBench and RoboDojo, the paper reports complementary benefits from Task and Action Knowledge, historical error correction with valid-knowledge retention, held-out improvement, and zero-shot transfer.These are the three stated contribution areas: experience-driven improvement under a fixed model and executor, evidence-driven HPK evolution, and systematic evaluation.Source: §1 · p. 3

2 RELATED WORK

Related work covers agentic manipulation harnesses and systems that evolve skills or other reusable components from experience. RoboHarn-Evo focuses on evolving physical knowledge for a fixed planner–executor system, with distinct checks for subtask completion and local action effects.

Harnesses for agentic manipulation

  • Agentic-manipulation systems span affordance-grounded skill selection and feedback replanning, VLM reasoning with VLA execution, and memory-guided frozen-VLA harnesses (Huang et al., 2026; Zhang et al., 2026).A systematic study identifies planner–executor orchestration as a determinant of hierarchical VLA performance.Source: §2 · p. 3

Self-evolution in robotics

  • Robotics self-evolution includes feedback-refined skills, hierarchical skill libraries, interaction-derived physical knowledge, procedural or runtime components, and policy refinement (Li et al., 2026a).RoboHarn-Evo checks subtask completion separately from local action effects to guide knowledge revision and retrieval eligibility, while preserving task goals during reuse.Source: §2 · p. 3

3 PROBLEM FORMULATION

The formulation treats robotic self-improvement as continual refinement of persistent knowledge from physical experience while the VLM and executor remain fixed. Episodes record observations and operations; a knowledge-update rule uses task progress and physical effects to revise what is available in subsequent episodes.

3 PROBLEM FORMULATION

  • The formulation defines the goal as continual refinement of persistent knowledge from physical experience under a fixed VLM and executor.Each episode uses the knowledge available before it, then updates the knowledge for subsequent episodes.Source: §3 · pp. 3–4
  • The overview separates an inner loop that executes operations and records outcomes from an outer loop that organizes experience into Task and Action Knowledge for later decisions.Concrete motions are computed from the current scene.Source: Figure 2 · p. 4
  • An episode records the instruction, observations, and robot-action sequences at operation boundaries, with retrieved knowledge conditioning tool-use decisions.The executor grounds parameterized tool calls in the current scene and provides execution feedback.Source: §3 · p. 4
  • The update rule revises the knowledge store from observed task progress and physical action effects, using episode observations, tool calls, and execution feedback.Held-out evaluation uses frozen knowledge snapshots on fixed configurations without feeding evaluation outcomes back into the store.Source: §3 · p. 4

4 METHOD

RoboHarn-Evo connects a fixed VLM to robot execution through inner HPK-guided manipulation and an outer loop that extracts and maintains knowledge from physical outcomes. Its hierarchy separates task choice and completion from action geometry and physical effects.

4.1 AGENTIC MANIPULATION HARNESS

  • Runtime memory and sequential task-then-action strategies guide grounded tool calls; effect checks distinguish physical operation success from completion of the enclosing subtask.Verified subtask completion advances task planning; otherwise, updated runtime state conditions the next decision.Source: §4.1 · pp. 4–5

4.2 HIERARCHICAL PHYSICAL KNOWLEDGE

  • Trajectory reflection creates Task entries for subtask applicability, goals, and completion, and linked Action entries for operation type, object-relative geometry, and observed effects.Skills group related entries and summarize when they are useful; experience changes knowledge content, applicability, and retrieval eligibility while the schema and execution primitives remain fixed.Source: §4.2 · p. 5

4.3 KNOWLEDGE RETRIEVAL AND MAINTENANCE

  • Task retrieval selects what to do, then conditioned Action retrieval selects how to act; evidence-driven maintenance revises knowledge and refreshes only affected Skill summaries.Task Knowledge is checked against subtask completion and Action Knowledge against intended physical effects; revisions become retrievable only when supported without applicable opposition.Source: §4.3 · pp. 5–6, Figure 3 · p. 5
  • Table 1 reports task success across six RMBench tasks, with Overall as their mean and same-model w/o HPK and Full HPK runs sharing an executor.Results other than X-VLA and Mem-0 use 20 held-out episodes per task; HarnessVLA is tested with GPT-5.5.Source: Table 1 · p. 6

5 EXPERIMENTS

The experiments test HPK’s contributions to manipulation, knowledge maintenance, continued-interaction learning, and transfer across benchmarks and real-world scenes. They evaluate six RMBench tasks and cross-benchmark transfer to RoboDojo.

5.1 EFFECTIVENESS OF HIERARCHICAL PHYSICAL KNOWLEDGE

  • 61.1 and 30.3 percentage-point overall-success gains over π0.5 and Harness VLA, respectively, averaged across three VLM backbones.Across six RMBench tasks, Full HPK also raises mean success over w/o HPK for Qwen3.8-27B, GPT-5.5, and GPT-6.Source: §5.1 · p. 6

5.2 KNOWLEDGE MAINTENANCE

  • At K = 80, GPT-5.5 and GPT-6 repair 41.7% and 50.0% of initial errors, respectively, deactivate another 41.7% each, and retain 95.8% of initially correct entries.Active-pool correctness reaches 88.9% with GPT-5.5 and 92.8% with GPT-6.Source: Figure 4 · p. 7, §5.2 · p. 8

5.3 SELF-IMPROVEMENT THROUGH CONTINUED INTERACTIONS

  • Held-out success rises from 48.3% to 75.0% with GPT-5.5 and from 70.0% to 88.3% with GPT-6 as interaction accumulates.At K = 80, RoboHarn-Evo exceeds Raw Memory by 16.7 and 15.0 percentage points and w/o Verification by 8.3 and 6.7 points for GPT-5.5 and GPT-6, respectively.Source: Figure 4 · p. 7, §5.3 · p. 8

5.4 KNOWLEDGE TRANSFER ACROSS ENVIRONMENTS

  • Transferred HPK raises mean RoboDojo success from 40.0% to 75.0% with GPT-5.5 and from 65.0% to 90.0% with GPT-6, without target-environment knowledge updates.Mean task progress also increases from 43.5 to 77.5 and from 68.5 to 91.5, respectively.Source: §5.4 · p. 8
  • Across three real-world tasks, RSI raises mean success from 41.7% to 62.5% and mean normalized task progress from 70.5% to 84.2%.Both metrics improve on every task; the VLM and executor remain fixed during updates.Source: Figure 5 · p. 9, §5.4 · p. 9
  • In the Press Button-to-Uncover, Count and Press case, the robot verifies the second green press before switching to blue, grounding contact geometry in the current scene.Source guidance is reused while HPK remains fixed during the target episode.Source: Figure 6 · p. 9, §5.4 · p. 9

6 CONCLUSION

RoboHarn-Evo evolves hierarchical physical knowledge from experience while keeping the VLM and low-level executor fixed. The experiments report improved knowledge quality, held-out performance, and transfer through reuse and revision of that knowledge.

6 CONCLUSION

  • Hierarchical Physical Knowledge links subtask decisions and completion conditions to object-relative action strategies and their physical effects.Physical feedback supports knowledge revision and reuse across episodes.Source: §6 · p. 10
  • Evidence-driven maintenance repairs or deactivates historical errors while preserving valid knowledge, and held-out performance improves as interaction accumulates.Task Knowledge and Action Knowledge contribute complementary benefits.Source: §6 · p. 10
  • Cross-benchmark and real-robot results show that simulation-derived knowledge supports transfer and can be refined through later physical interaction.The reported approach improves manipulation without updating model parameters.Source: §6 · p. 10

A.1 SIMULATION EXPERIMENTAL SETUP

The setup evaluates manipulation across RMBench and RoboDojo, with separate protocols for simulation tasks, knowledge transfer, and real-world evaluation. Task completion and partial progress are scored under ordered subtask requirements.

Benchmarks and tasks

  • RMBench is the main platform, with six tasks spanning planning, action grounding, and their interaction; transfer uses two RoboDojo tasks and ten episodes per task.Knowledge correction and self-evolution focus on Swap Blocks and Put Back Block.Source: §A.1 · p. 15

Real-world hardware and evaluation

  • The real-world platform is an AgileX Cobot Magic with two follower arms and one front-view plus two wrist-mounted RGB-D cameras.The three evaluated tasks use the same real-world comparison protocol.Source: Figure 7 · p. 15, §A.1 · p. 16

π0.5 baseline training and deployment

  • The π0.5 baseline is fine-tuned on 450 RMBench simulation demonstrations, predicts 50-step action chunks, and is deployed without real-world fine-tuning.The final checkpoint is deployed directly, with policy parameters fixed during evaluation.Source: §A.1 · p. 16

Evaluation Rule

  • The real-world comparison includes π0.5 and RoboHarn-Evo with simulation-derived HPK or HPK updated through real-world interaction, evaluated on eight scenes per task.It reports full-task success and normalized task progress.Source: §A.1 · p. 16

Task progress and success

  • Subtasks score 0, 0.5, or 1; credit follows task order, and success requires full task score with all instruction constraints satisfied.Aggregated success averages task-level percentages with equal weight.Source: §A.1 · p. 16

Task requirements and scoring criteria

  • Tasks require ordered cup-block coverings, instruction-mapped digit button presses followed by a remaining-button press, or uncovering and color-count presses.The third task assigns cup removal and button pressing sequentially to different arms.Source: §A.1 · pp. 16–17

B EXAMPLES OF PERSISTENT TASK AND ACTION KNOWLEDGE

The appendix presents a representative persistent knowledge store for Cover Blocks and describes how its task-level and action-level entries are organized.

B EXAMPLES OF PERSISTENT TASK AND ACTION KNOWLEDGE

  • The source package comes from ten fixed demonstration trajectories and contains six Task Knowledge entries grouped into two Task Skills and four Action Knowledge entries grouped into four Action Skills.Identifiers and runtime bindings are omitted because they are not part of the VLM-facing representation.Source: B EXAMPLES OF PERSISTENT TASK AND… · p. 18
  • Table 4 summarizes the complete persistent store, while Tables 5 and 6 present one Task Knowledge entry and its complementary Action Knowledge entry.Figure 9 grounds both entries in the corresponding execution sequence.Source: B EXAMPLES OF PERSISTENT TASK AND… · p. 18
  • The examples preserve the semantic fields stored by the method, rather than identifiers or runtime bindings.Those omitted fields are not part of the VLM-facing knowledge representation.Source: B EXAMPLES OF PERSISTENT TASK AND… · p. 18

B.1 TASK KNOWLEDGE EXAMPLE

The Cover Blocks example links Task Knowledge for choosing the next subtask with Action Knowledge for carrying it out. Its observations illustrate how an action effect can be checked separately from subtask completion.

B.1 TASK KNOWLEDGE EXAMPLE

  • Task Knowledge selects the next block by requested color order rather than spatial position, and the example shows red after the selected lid is moved.Table 4 catalogs the extracted Cover Blocks entries, while Table 5 presents a representative Task Knowledge entry.Source: §B.1 · p. 18, Table 4 · p. 19, Table 5 · p. 19

B.2 ACTION KNOWLEDGE EXAMPLE

  • Action Knowledge distinguishes removing an enclosing lid from grasping a free-standing lid and uses the visible red block with the moving lid as an expected-effect check.The example’s frames trace the effect and subsequent set-aside completion in one execution, not repeated validation.Source: §B.2 · p. 19
  • These observations are illustrative evidence from one fixed demonstration, not an aggregate reliability evaluation or evidence of real-robot effectiveness.The frames show a state transition from the applicable condition to the expected effect and subtask completion.Source: §B.2 · p. 19

C ADDITIONAL METHOD DETAILS

RoboHarn-Evo evolves the knowledge guiding a fixed VLM and executor through an interaction–update cycle. The section’s example materials include an Action Knowledge entry and frames associated with two representative entries.

C ADDITIONAL METHOD DETAILS

  • RoboHarn-Evo couples execution with persistent knowledge evolution while keeping the VLM and robot executor fixed.Task Knowledge specifies the subtask objective, Action Knowledge specifies how to act toward it, and physical evidence informs revision before reuse.Source: C ADDITIONAL METHOD DETAILS · p. 19
  • Table 6 presents a representative Action Knowledge entry, whose applicability and expected-effect criteria guide how the selected subtask is realized.The entry distinguishes removing an enclosing lid from grasping a free-standing lid and gives an observable effect for verification.Source: Table 6
  • Figure 9 associates Task Knowledge evidence from frames 512–687 with Action Knowledge evidence for grasping the enclosing lid from frames 512–600.The final frame also verifies completion of the subsequent set-aside action.Source: Figure 9 · p. 20

C.1 HIERARCHICAL KNOWLEDGE REPRESENTATION

The method represents manipulation knowledge at task and action levels, retrieves it under the current scene and goal, and revises it using physical evidence. Its episode loop preserves distinctions between operation effects, subtask completion, and evidence supporting reusable knowledge.

C.1 HIERARCHICAL KNOWLEDGE REPRESENTATION

  • Runtime memory tracks the active subtask, objects, relations, operations, and outcomes, while each episode’s knowledge package preserves subtask order and operation-linked observations.Observed facts and inferred rationales are recorded separately; atomic entries retain shared task context and temporal associations.Source: §C.1 · pp. 20–21
  • Task entries encode when a subtask applies, its purpose, selection rationale, and completion condition; Action entries encode operation, object relations, geometry, and intended physical effect.Geometric strategies are instantiated from current observations rather than stored robot poses, and the entries are grouped into Skills by related decision context.Source: §C.1 · pp. 20–21

C.2 HIERARCHICAL RETRIEVAL AND GOAL-CONSISTENT EXECUTION

  • Task retrieval refines the planner’s subtask using instruction, runtime memory, and matching conditions, while Action retrieval selects applicable geometry for that chosen subtask and operation.If no Task entry applies, the baseline planner decision remains; Action Knowledge guides realization rather than changing the target.Source: §C.2 · p. 21
  • Scene grounding prioritizes retrieved geometry only among feasible candidates consistent with the task goal, and the agent re-observes or replans if none is admissible.For larger stores, Skill routing shortlists entries before final selection; exhaustive recall is used for stores of at most 24 entries.Source: §C.2 · pp. 21–22

C.3 PHYSICAL EVIDENCE AND KNOWLEDGE REVISION

  • Evidence is attributed to a specific entry version and source execution, with support, opposition, or unverified labels; out-of-scope executions do not count for that entry.Retrieval considers supported entries whose conditions apply, while unverified outcomes do not increase support or opposition.Source: §C.3 · p. 22
  • Operation effects and subtask completion are assessed separately, so a successful grasp can support Action Knowledge while the enclosing Task entry still awaits its completion condition.Delayed observations can refine the pending judgment for the same attempt rather than count as a second trial.Source: §C.3 · p. 22, §C.2 · p. 22

C.4 THE INTERACTION–UPDATE CYCLE

  • Equivalent entries are consolidated without merging distinct interactions, and Algorithm 1 feeds each episode’s execution evidence into persistent updates for subsequent trajectories.The cycle assesses new entries, reviews historical ones, checks revisions, and retains the updated store while the VLM and executor remain fixed.Source: §C.3 · p. 22, §C.4 · p. 23

C.3 PHYSICAL EVIDENCE AND KNOWLEDGE REVISION

  • Maintenance revises challenged content against executions that tested the revised condition or strategy, preserving evidence by version and making only evidence-supported revisions retrievable.Withdrawal prevents reuse of a challenged claim; repair establishes a supported replacement, and affected Skill summaries are refreshed for subsequent episodes.Source: §C.3 · p. 23

D.1.1 METRIC DEFINITIONS

This section defines knowledge-quality and RoboDojo transfer metrics and describes runtime skill interfaces for perception, memory, planning, tool use, and verification. These skills ground decisions in current physical evidence, update subtask status, and guide whether to continue, retry, or replan while preserving uncertainty.

D.1.1 METRIC DEFINITIONS

  • Reviewed-pool correctness, error resolution, and valid-knowledge retention are tracked separately; unverified reviewed entries stay in the denominator, while unaudited entries are excluded.Sampled reviews require the sampling design and coverage, with weighted pool estimates for unequal-probability sampling. Error resolution requires independently verified repairs or inactivation without an erroneous active descendant.Source: §D.1.1 · p. 24

D.2 Q4: FROZEN-SOURCE TRANSFER AND TARGET ADAPTATION

  • RoboDojo transfer uses 10 episodes per condition; partial-progress scores range from 0 to 1.0 by verified stage, while terminal-only tasks are binary, so mean Score equals Success Rate.The intermediate scores are 0.05, 0.15, and 0.30; full-sequence completion and return of both arms earns 1.0.Source: §D.2 · p. 24

E.1 OVERVIEW AND ORGANIZATION

  • The runtime library contains 28 prompt-level skills across six categories, separate from learned policies, model parameters, and AFK’s semantic Skills.The categories are Perception, Memory, Planning, Monitoring, Tool Calling, and Effect Verification; experience records are maintained separately from runtime prompts.Source: §E.1 · pp. 24–25

E.2 PERCEPTION SKILLS

  • Perception Query Normalization standardizes object, role, relation, and identity queries by binding phase, preserving reference sets and requiring exact advertised identities for execution-facing selection.When a previous query set is rejected, the skill uses the failure reason to correct it rather than repeating it unchanged.Source: §E.2 · pp. 25–26

E.3 MEMORY SKILLS

  • Observation Memory Summarization writes committed facts and a recent execution digest into existing memory surfaces, updating progress only from supported evidence and preserving uncertainty.It retains task-relevant execution consequences and uses runtime position validity—not visibility alone—to determine whether tracked-object geometry remains usable.Source: §E.3 · pp. 27–28

E.4 PLANNING SKILLS

  • The Planning category contains five skills spanning task planning, runtime-state interpretation, monitored execution, and data-collection workflows.The paper provides two representative examples rather than describing every planning skill individually.Source: §E.4 · p. 28, §E.4.1 · p. 29

E.4.1 PLANNING SKILL INVENTORY

  • Table 9 enumerates the complete planning skill set, including specialized execution and data-collection workflows.The following discussion focuses on two representative skills that define the core planning interface.Source: Table 9 · p. 28, §E.4.1 · p. 29

E.4.2 REPRESENTATIVE PLANNING SKILLS

  • Control Turn Planner selects an evidence-backed executor subtask, while Runtime State Reasoning treats structured runtime fields as authoritative for control decisions.The planner may finish only when the runtime’s global-success signal is true; geometry marked current-verified supports grounded action, memory-valid geometry permits only a safe approach, and motion-uncertain historical coordinates are unusable.Source: §E.4.2 · pp. 29–30
  • Fresh structured state governs planning: preserve verified roles and memory, reject motion-uncertain coordinates, and restrict safe-motion repair while geometry is pending.The planning stack reads task, scene, monitor, manipulation, and recovery state before producing the next control turn.Source: §E.4.2 · p. 31

E.5 MONITORING SKILLS

  • OOD monitoring classifies execution health relative to the active subtask, returning one evidence-based scenario and conservatively choosing none when evidence is insufficient.It classifies rather than executes recovery, and does not replace progress or task-success detection.Source: §E.5 · pp. 32–33

E.6 TOOL CALLING SKILLS

  • The Tool Calling category comprises 19 skills: one router, ten reusable primitives, and eight workflows that construct situation-dependent sequences.Tables 10 and 11 enumerate the skill set; the text selects representative examples across routing, primitive, and workflow levels.Source: §E.6 · p. 33, §E.6.1 · p. 33

E.6.1 TOOL-CALLING SKILL INVENTORY

  • Tables 10 and 11 enumerate the complete Tool Calling skill set, while the inventory discussion presents three representative examples rather than every skill contract.The examples span routing, primitive, and workflow levels.Source: §E.6.1 · p. 33

E.6.2 REPRESENTATIVE TOOL-CALLING SKILLS

  • The router selects one available workflow from the explicit monitor signal and execution context, preferring least-disruptive recovery and choosing retry, replan, or abort intent.It does not reclassify observations or generate tool calls; runtime validates the selected workflow before invocation.Source: §E.6.2 · pp. 33–34
  • Grounded-motion calls use public scene or target identities and runtime-resolved geometry; the grasp-loss workflow composes available tools for the observed state and returns retry, replan, or abort.The motion primitive fails closed when required grounding or authority is unavailable, while recovery sequences are situation-specific rather than mandatory scripts.Source: §E.6.2 · pp. 35–37

E.7 EFFECT VERIFICATION SKILLS

  • After tool execution, effect verification compares fresh before-and-after evidence to determine whether the physical effect supports the current subtask transition.For displacement, it uses current observed positions; history-smoothed positions support identity continuity rather than measuring the latest action’s movement.Source: §E.7 · p. 38
  • Effect verification classifies outcomes as verified, contradicted, or unverified and uses fresh physical evidence to determine subtask status and the next control recommendation.For grasp or carry, it compares object and end-effector motion; placement requires a verified runtime validator result; transient contact requires a complete grounded contact-and-release cycle from the same arm.Source: §E.7 · p. 38
  • Verification follows action execution and fresh post-action evidence, updates local subtask state, and supplies a concise action-oriented memory update that may feed AFK assessment.The runtime uses the result to continue, retry, or return to planning; AFK separately assesses whether the execution event supports, opposes, or leaves a persistent knowledge entry unverified.Source: §E.7 · pp. 38–39
  • Successful tool execution, gripper commands, planner intent, and historical poses do not establish physical effects; current effects require before-and-after grounding.Placement is verified only when the runtime validator reports verified=true; a successful final move or gripper-open command is insufficient.Source: §E.7 · p. 39
  • A verified transient contact cycle counts as one event, while completion requires fresh evidence of the current subtask’s physical stop condition.Missing, occluded, contradictory, or insufficient evidence remains unverified; the global environment-success signal remains authoritative for task completion.Source: §E.7 · p. 39

Related papers

Loading 2609.37583v2…