PaperlayerView exact revision on arXiv

Source-linked AI summary

Skill2Real: Agentic Skill Learning for Zero-Shot Sim-to-Real Robot Manipulation

Xincheng He, Siyu Ma, Chang Yu, Yunuo Chen, Yanjia Huang, Ying Nian Wu, Yin Yang, Chenfanfu Jiang

arXiv:2610.02788v1cs.RO

TL;DR

Simulation-learned robot skills can fail to transfer across differences in perception, dynamics, and embodiment. Skill2Real learns a frozen hierarchy through a shared interface, with Astra’s unseen LIBERO-Pro Long success rising from 2.0% to 56.3%.

  • Problem

    Long-horizon manipulation requires reliable local skills, task-level composition, and recovery despite simulation-to-reality gaps in perception, dynamics, contact, and embodiment.

    Source: §1 · p. 2

  • Method

    Skill2Real learns a two-level hierarchy with a Proposer–Verifier–Governor loop, using privileged simulation evidence for validation while grounding frozen skills through public observations and a shared interface.

    Source: §1 · p. 2

  • Results

    Frozen hierarchical skills generalize across tasks, scenes, and the real robot; Astra’s unseen LIBERO-Pro Long success rises from 2.0% to 56.3%.

    Source: §6 · p. 9

  • Takeaways & Limitations

    The results support transferring simulation-learned programs across tasks and to the real robot while keeping the learned hierarchy fixed.

    Source: §6 · p. 9

    The current pure code-as-policy interface relies on a robot backend for perception and low-level control, and future work targets lower VLM inference latency.

    Source: §6 · p. 9

Abstract

from arXiv · show

Transferring robotic skills from simulation to reality requires task knowledge that remains usable across differences in perception, dynamics, and embodiment. We introduce Skill2Real, an agentic policy framework that learns executable skills through a shared application programming interface (API). A Proposer-Verifier-Governor (PVG) loop uses privileged simulation evidence to diagnose outcomes and validate updates, while keeping learned skills grounded in public observations and API semantics. The Cerebellum first acquires local manipulation skills; the Brain then learns task-level composition with the Cerebellum frozen. Both memories transfer to the real robot without task-policy fine-tuning or skill-memory updates. As GPT-5.6 Sol learns skills on LIBERO-90, evaluating each frozen checkpoint with GPT-6 Astra raises LIBERO-Pro Long success from 2.0% to 56.3%, without training on Pro Long. Independent Robosuite training reaches 85.1% and 89.4% mean success with Sol and Opus 5 across seven tasks, respectively. Frozen Sol-trained LIBERO-90 skills achieve 78.75% mean completion across four real-world manipulation tasks with Astra. Removing the Verifier or Governor during LIBERO-90 training lowers final Pro Long success by 17.3 and 13.3 percentage points, respectively. These results support learning and transferring a hierarchy of executable skills through a common robot interface.

1 INTRODUCTION

Long-horizon sim-to-real manipulation requires reusable local skills, task-level composition, and recovery despite gaps in perception, dynamics, and embodiment. Skill2Real learns a frozen hierarchy through a shared interface, with simulation evidence used for validation and public observations grounding execution.

1 INTRODUCTION

  • Existing approaches leave a bottleneck: acquiring reusable, composable task knowledge from privileged simulation evidence without spending learning on task-agnostic low-level representations.The introduction contrasts this gap with methods that adapt policies or representations, code-generation from supplied primitives, and skill libraries accumulated through human guidance or real-world feedback.Source: §1 · p. 2
  • Skill2Real separates privileged simulation supervision from execution: a shared code-based interface grounds transferable task knowledge in public observations and backend-specific robot operations.The framework transfers executable behavior without incorporating simulator-only state into the deployed policy or requiring real-world task-policy adaptation.Source: §1 · p. 2
  • The PVG loop uses privileged outcomes for public-facing diagnoses and validation-gated memory updates, while the Cerebellum learns local skills before the frozen library is composed by the Brain.The hierarchy supports task-level planning and recovery.Source: §1 · p. 2
  • Astra’s unseen LIBERO-Pro Long success rises from 2.0% to 56.3%, while UR5e completion reaches 78.75% with the full hierarchy versus 56.25% with local skills alone.The evaluations test frozen skill libraries across task suites, Proposers, and the real robot.Source: §1 · p. 2

2 RELATED WORK

Related work spans sim-to-real adaptation, language-agent code policies, and hierarchical skill learning. Skill2Real differs by transferring validated executable task knowledge through a shared interface, using a two-level hierarchy of local skills and task programs.

2 RELATED WORK

  • Sim-to-real methods primarily address visual, dynamics, and execution gaps by learning policies or representations invariant to domain shift; Skill2Real instead transfers validated executable task knowledge.Foundation VLMs provide cross-domain reasoning, while the shared interface leaves perception, calibration, and low-level execution to the backend.Source: §2 · p. 3
  • Code-policy and agentic systems generate or improve robot programs, while Skill2Real assigns proposal, privileged-outcome diagnosis, and memory admission to distinct agents.Its agents learn local skills and task programs as frozen, publicly grounded knowledge.Source: §2 · p. 3
  • Unlike hierarchies whose transfer units are typically policies, latent skills, fixed stages, or episodic records, Skill2Real freezes reusable local skills and learns task-level programs that recombine them.The shared interface supports reuse across tasks, scenes, and deployment domains.Source: §2 · p. 3

3 METHOD

Skill2Real learns executable robot skills in simulation through a shared interface, separating privileged learning-time supervision from public information used to execute skills. Its hierarchy moves from reusable local manipulation skills to task-level compositions, then transfers the frozen memories to the real robot.

3.1 SHARED CODE-BASED POLICY INTERFACE

  • The shared API gives simulated and real backends common operation semantics while leaving perception, calibration, and low-level control environment-specific.Learned knowledge transfers as executable programs over the interface, not as model parameters.Source: §3 · p. 3, §3.1 · p. 4

3.2 MULTI-AGENT LEARNING FRAMEWORK

  • The Proposer uses public observations and API outcomes, while the Verifier turns privileged simulation evidence into public-only feedback and the Governor admits updates supported across validation rollouts.Only the Proposer, shared interface, and frozen skill memories remain at deployment.Source: §3.2 · p. 5

3.3 HIERARCHICAL SKILL TRAINING

  • The Cerebellum stores local manipulation skills, and the Brain composes them into long-horizon task behavior using the frozen Cerebellum library as context.Both levels use the PVG procedure, but differ in representation granularity and training context.Source: §3.3 · pp. 5–6

3.4 ZERO-SHOT SIM-TO-REAL DEPLOYMENT

  • After simulation training, the real robot grounds both frozen memories through real observations and the shared API semantics.The transferred hierarchy consists of executable Cerebellum and Brain skills rather than model parameters.Source: §3 · p. 3, §3.4 · p. 6

4 EXPERIMENTAL SETUP

The experiments train separate skill hierarchies on LIBERO-90 and Robosuite, then evaluate frozen checkpoints on unseen long-horizon tasks and in real-world manipulation. Evaluation compares the base Proposer, Cerebellum-only skills, and the full hierarchy, with a Brain-only control on the real robot.

4.1 SIMULATION SETUP

  • Separate Sol and Opus 5 hierarchies are trained from scratch on LIBERO-90 and seven Robosuite tasks, with a Franka Panda in LIBERO-90 and single- and two-arm Robosuite settings.The Proposer requests configured RGB-D camera views through the shared interface.Source: §4.1 · p. 6

4.2 TRAINING SETUP

  • Each iteration covers every task with five seeded training rollouts per task, and Governor runs assess candidates over 25 validation rollouts with disjoint seeds.Across C1–B4, this totals 3,150 LIBERO-90 episodes and 245 Robosuite episodes; Figure 6 shows success rising through both hierarchy stages in both settings.Source: §4.2 · p. 6, Figure 6 · p. 7

4.3 REAL-WORLD SETUP

  • Real-world evaluation uses a single UR5e arm with a Pika gripper and RGB-D cameras across pick-and-place, sorting, equation assembly, and drawer manipulation.Deployment uses no real-world task demonstrations, task-policy fine-tuning, or online skill-memory updates.Source: §4.3 · p. 7

4.4 EVALUATION SETUP

  • Evaluation compares the base Proposer, Cerebellum-only skills, and the full hierarchy, with a Brain-only real-robot control; real tasks use 20 trials per method.Pro Long checkpoints are evaluated across 20 tasks and 10 seeds; the fixed C3/B3 comparison uses 20 trials per task under each perturbation setting.Source: §4.4 · p. 7

5 RESULTS

The results evaluate transfer from learned skill hierarchies to unseen tasks, real-world manipulation, and new Proposers, and test the Verifier and Governor’s roles. Across these settings, the experiments report gains from local skills and task-level composition, with a position-perturbation weakness against Zetta.

5 RESULTS

  • The evaluation compares checkpoint transfer, independent Robosuite learning, real-world transfer, cross-Proposer generalization, and PVG-role ablations.The results are organized around the contribution of each hierarchy level, transfer breadth, and the Verifier and Governor’s contributions.Source: §5 · p. 7

5.3 ZERO-SHOT LONG-HORIZON GENERALIZATION

  • With Astra, Skill2Real reaches 55.3% Overall and 75.0% Task on LIBERO-Pro Long, beating Zetta’s 51.5% and 63.0%; with Opus 4.6 it reaches 35.3% Overall, above ASPIRE’s 30.5%.The Astra evaluation uses the fixed Sol-trained C3/B3 hierarchy.Source: Figure 7 · p. 8, §5.3 · p. 9

5.1 SKILL TRANSFER ACROSS TRAINING CHECKPOINTS

  • On unseen LIBERO-Pro Long, Cerebellum learning raises Overall success from 2.0% to 35.3% for Astra and from 0.5% to 29.3% for Opus 5; Brain learning adds 21.0 and 19.8 percentage points.The Brain is trained with the C3 Cerebellum fixed, reaching 56.3% for Astra and 49.0% for Opus 5.Source: §5.1 · p. 8
  • Across seven Robosuite tasks, Brain composition lifts mean success to 85.1% for Sol and 89.4% for Opus after Cerebellum learning reaches 45.6% and 49.6%.Brain learning improves all seven tasks.Source: §5.1 · p. 8

5.2 ZERO-SHOT REAL-WORLD TRANSFER

  • 78.75% mean real-world completion with the full hierarchy exceeds 27.50% without learned skills, Brain-only at 37.50%, and Cerebellum-only at 56.25%.The mean averages R1–R4; the full hierarchy is the only configuration to complete drawer manipulation, at 50%.Source: Table 1 · p. 8, §5.2 · p. 8

5.4 VERIFIER AND GOVERNOR ABLATIONS

  • 56.3% Overall with full Skill2Real exceeds 39.0% without the Verifier and 43.0% without the Governor, losses of 17.3 and 13.3 percentage points.The gap appears during Cerebellum learning and persists as the Brain composes the fixed C3 library.Source: §5.4 · p. 9

6 CONCLUSION AND DISCUSSION

Skill2Real learns a two-level hierarchy through privileged simulation evidence and a shared code-based policy interface. The paper reports transfer with the hierarchy fixed and identifies interface and inference-latency extensions as future work.

6 CONCLUSION AND DISCUSSION

  • The Cerebellum learns reusable local skills and the Brain composes them into long-horizon programs, with both levels expressed through a shared interface grounded in real observations.A foundation VLM grounds the frozen hierarchy during deployment.Source: §6 · p. 9
  • Across LIBERO-90, Robosuite, LIBERO-Pro Long, and UR5e, the two levels provide complementary capabilities; the frozen hierarchy reaches 78.75% mean UR5e completion versus 56.25% for the Cerebellum alone.The Proposer–Verifier–Governor ablations also show that privileged diagnosis and evidence-based memory admission materially improve skill acquisition.Source: §6 · p. 9
  • The current shared policy interface relies on a robot backend for perception and low-level control; future work would expose end-to-end policies as callable tools and reduce VLM inference latency.Source: §6 · p. 9

AI DISCLOSURE

The authors disclose generative AI assistance with manuscript drafting and language polishing, and state that they reviewed and revised the AI-assisted text.

AI DISCLOSURE

  • Generative AI assisted with drafting and language polishing; the authors reviewed and revised the resulting text and accept responsibility for the final content.Source: AI DISCLOSURE · p. 10

A REAL-WORLD TASKS AND VISUAL GENERALIZATION

The appendix describes four real-world manipulation tasks and provides qualitative examples of reuse across tabletop appearances and lighting conditions. The dim-light example documents execution but not final equation acceptance.

A REAL-WORLD TASKS AND VISUAL GENERALIZATION

  • The real-world task suite uses R1–R4 for pick-and-place, sorting, equation assembly, and drawer manipulation, with object instances specified per trial.Figure A.1 shows one recorded scene for each task.Source: Table A.1 · p. 15, Figure A.1 · p. 15

A.1 TABLETOP APPEARANCE AND LIGHTING VARIATION

  • Recorded deployments show the potato-to-plate operation on white and floral surfaces and three placements into color-labeled boxes on red gingham.Reviews confirm completion of these examples; they complement aggregate task-completion results.Source: §A.1 · p. 16, Figure A.2 · p. 16
  • Under dim, uneven side lighting, the robot performs two answer-cube pick-and-place sequences, but the record ends before final equation acceptance.This qualitative example supplies no separate condition-wise success rate and adds no trials to Table 1.Source: §A.1 · p. 16, Figure A.2 · p. 16

B.1 SHARED CONFIGURATION AND EVALUATION PROTOCOL

LIBERO-90 evaluation follows fixed memory stages and held-out protocols, with the Brain trained after the Cerebellum is fixed. Both models improve across stages, though family-level gains are uneven and reported stage comparisons do not isolate hierarchy from additional training.

B.1 SHARED CONFIGURATION AND EVALUATION PROTOCOL

  • C0 is the no-skill baseline; C1–C3 train the Cerebellum, then B1–B4 train the Brain with the C3 Cerebellum held fixed.Sol and Opus train separate libraries; both memories are frozen during evaluation.Source: §B.1 · pp. 16–17, Figure B.1 · p. 18
  • Stage comparisons combine additional training with the second memory level, rather than isolating hierarchy at a matched training budget.Reported success rates are point estimates for fixed libraries and do not measure variability across independent training runs.Source: §B.1 · p. 17

B.2 LIBERO-90 FAMILIES AND STAGE COMPARISONS

  • Each LIBERO-90 evaluation uses five held-out initial states per task, or 450 trials; suite scores average across all 90 tasks.Family scores instead use unweighted means of their task success rates; the six family denominators sum to 450 and are not task counts.Source: §B.2 · p. 17, Table B.1 · p. 18
  • Opus reaches 100.0% in caddy storage at B4, up from 56.7% at C3, while its stacking score remains 40.0%; family gains are not uniform.The family panels use independent vertical scales; the additional Brain stage brings larger gains in caddy storage, basket/tray storage, and surface placement than in stacking for both models.Source: §B.2 · p. 17
  • 62.0% at B4 for Sol (from 14.2% at C0) and 73.3% for Opus (from 18.7% at C0); Brain learning adds 23.8 and 29.6 percentage points, respectively.The source-suite curves evaluate independently trained libraries on LIBERO-90.Source: §B.2 · p. 18

B.3 INDEPENDENT ROBOSUITE TRAINING

Skill2Real is trained anew on Robosuite, with the C3 Cerebellum fixed during Brain learning and both memories frozen for evaluation. Brain-stage gains span assembly and bimanual tasks, while the final task leader varies.

B.3 INDEPENDENT ROBOSUITE TRAINING

  • Libraries are trained anew on Robosuite rather than initialized from LIBERO-90, and the experiment broadens task-suite coverage within simulation, not to a different physics engine.Mean success gives equal weight to the seven reported tasks; three configured pick-and-place tasks have no reported results and are excluded.Source: §B.3 · pp. 18–19
  • From C3 to B4, nut assembly improves by 62 and 50 percentage points for Sol and Opus, two-arm lifting by 48 and 46 points, and handover by 52 and 60 points.Cube lifting rises from 94% to 98% for Sol and from 77% to 95% for Opus.Source: §B.3 · p. 19
  • At B4, Opus leads on five of seven tasks, while Sol scores higher on cube lifting (98% versus 95%) and spill wiping (100% versus 99%).Source: §B.3 · p. 19

B.4 LONG-HORIZON TRANSFER PROTOCOL AND PER-TASK RESULTS

The long-horizon evaluation tests frozen LIBERO-90 skills on unseen LIBERO-Pro Long tasks under position and instruction perturbations. The paper also describes a separate real-robot evaluation using frozen simulation-trained memories and no task-policy fine-tuning or skill-memory updates.

B.4 LONG-HORIZON TRANSFER PROTOCOL AND PER-TASK RESULTS

  • LIBERO-Pro Long uses LIBERO-90-trained, frozen memories; target tasks contribute no skill construction, prompt examples, debugging, bootstrapping, or memory updates.Overall success is the equal mean of position-perturbation and instruction-perturbation success rates.Source: §B.4 · p. 19
  • Under instruction perturbations, Astra scores 90.0% on two-mug and mug-plus-pudding, but 35.0% and 25.0% under position perturbations; neither Proposer succeeds on soup-plus-cream-cheese under position perturbations.Position perturbations are harder in aggregate, but the Pos–Task gap varies across tasks: Astra scores 80.0% on Pos and 30.0% on Task for soup-plus-tomato-sauce.Source: §B.4 · p. 20

B.5 REAL-ROBOT EVALUATION DETAILS

  • Real-robot evaluation covers four UR5e tasks, with 20 trials per method on each; both memories are learned in simulation and frozen, without real-world task demonstrations, fine-tuning, or online skill-memory updates.The comparison includes the full hierarchy and single-memory conditions using the same test-time model and robot interface.Source: §B.5 · p. 22

B.6 PVG TRAINING ABLATIONS AND TRANSFER EVALUATION

Removing either the Verifier or Governor lowers final LIBERO-Pro Long success, and the full PVG system also has the highest recorded LIBERO-90 success. The appendix details how a shared public API supports execution and how PVG diagnoses failures and admits supported skill updates.

B.6 PVG TRAINING ABLATIONS AND TRANSFER EVALUATION

  • At B4, removing the Verifier or Governor reduced GPT-6 Astra LIBERO-Pro Long Overall success from Full Skill2Real’s 56.25% to 39.0% and 43.0%, losses of 17.25 and 13.25 percentage points.The removed roles affect LIBERO-90 training; neither participates in target evaluation.Source: §B.6 · p. 22
  • On LIBERO-90, Full Skill2Real reaches 62.0% at B4 versus 38.0% without the Verifier and 40.0% without the Governor; Figure B.4 tracks the variants across C0–B4.All variants start at C0 with 64/450 (14.2%), and the evaluations use five initial states per task.Source: §B.6 · p. 22, Figure B.4 · p. 23

C SHARED ROBOT API AND EXECUTION CONTRACT

  • The shared API defines an 18-operation executable interface; skills compose its operations, while perception exposes measurements and excludes simulator handles, ground-truth poses, hidden scene structure, and success predicates.The real backend may use different sensing implementations while sharing API semantics.Source: C SHARED ROBOT API AND EXECUTION… · p. 23, §C.1 · p. 23

C.2 EXECUTION STATE, POSE CONSTRUCTION, AND CONTROL

  • Runtime calls bind to the current execution step, pose construction uses carry geometry, and only the final four of the ten state, compilation, and control operations issue arm or gripper commands.Pose construction does not itself move the robot.Source: §C.2 · pp. 23, 25, Table C.2 · p. 24

C.3 FROM AN OBSERVED TARGET TO AN EXECUTED PLACEMENT

  • Placement proceeds from calibrated RGB/depth and proposer-chosen target binding through pose compilation and checked motion commands to fresh post-release evidence; compilation or command completion alone does not establish task success.The target binding and carry geometry are used to construct support-relative release and clearance poses.Source: §C.3 · p. 25

C.4 RELATION TO PVG LEARNING AND FROZEN DEPLOYMENT

  • PVG keeps privileged simulation evaluation outside the robot API: the Verifier converts evaluator evidence into public feedback, the Governor validates updates, and deployment uses frozen memories without either role.Interface-specification checks track a contract version but do not establish physical success or equivalent perception and control across backends.Source: §C.4 · p. 25

D.1 FROM EXECUTION FAILURE TO A CANDIDATE SKILL

  • In the recorded bowl-transfer case, failed lifts prompted a Verifier diagnosis of contact and retention, followed by a Proposer contact revision toward the bowl’s narrower lower profile and a short-lift check.Subsequent successful executions supported review of candidate entries.Source: §D.1 · p. 26, D RECORDED PVG LEARNING AND ADMISSION · p. 26

D.2 WHY ONE CANDIDATE WAS ADMITTED AND THREE WERE REJECTED

  • The Governor admitted a Brain rule selecting the distinct bowl with the leftmost image center, conditional on viewpoint, visibility, or occlusion changes; local-contact and untested progress-checking proposals stayed out.Whole-program success supported review but did not validate every proposal or isolate the admitted rule’s causal contribution.Source: §D.2 · p. 26

E PVG SKILL-MEMORY TEMPLATES AND EXAMPLES

This appendix presents templates and examples showing how Skill2Real represents task-level and local skills and operationalizes its PVG workflow through a shared API, with public observations and evidence guiding execution, diagnosis, proposal, validation, and admission. It clarifies memory boundaries and checks; evaluation and deployment use frozen memories without skill-memory updates, and the examples are not additional evaluated rollouts.

E.1 INITIAL BRAIN TEMPLATE

  • Brain memory composes objectives and binds public inputs, while frozen cerebellum skills handle local actions; filtered Verifier feedback and Governor validation govern memory updates.The initial brain template has no learned task-level entries, but retains the API documentation and writing scaffold.Source: §E.1 · p. 27

E.2 ILLUSTRATIVE SKILL REPRESENTATION UNDER THE SHARED API

  • An illustrative local grasp skill uses public geometry to select rim contact, then checks contact and retention after a short lift rather than treating command completion as success.Its representation separates persistent guidance from episode-specific bindings and does not establish performance beyond its supported scope.Source: §E.2 · p. 28

E.3 DOCUMENTED CALLS AND THE PLACEMENT HANDOFF

  • The shared-API sequence binds masks and measurements to the matching observation and depth, inspects results before dependent calls, and passes runtime carry geometry to placement.Missing carry geometry or an invalid binding leaves placement unresolved rather than prompting invented state.Source: §E.3 · p. 29

E.4 CEREBELLUM SKILLS BEYOND GRASPING

  • For receptacle openings, the cerebellum compares candidate regions with the parent using area and overlap or containment, leaving localization unresolved when public evidence is ambiguous.The selected mask is measured against depth and calibration from the same observation; unsupported slivers are not treated as actionable openings.Source: §E.4 · pp. 29–30
  • For oblique receptacle views, the local placement skill selects a supported interior target from visible bounds and checks the released object’s relation instead of trusting its apparent center.Placement uses runtime-held-object geometry and bound observation references; a completed descent alone does not establish containment.Source: §E.4 · p. 30

E.5 A FINAL BRAIN SKILL FOR RELATIONAL COMPOSITION

  • The final brain example places the support bowl in the tray before stacking only when instruction permits, then rebinds the moved support using fresh containment evidence.If the support cannot be identified unambiguously or is outside the tray, the dependent placement is withheld; the example is not an isolated causal evaluation.Source: §E.5 · pp. 30–31

E.6 TASK PROGRESS AND LOCAL RECOVERY IN DRAWER MANIPULATION

  • In drawer tasks, Brain memory permits closure only after fresh evidence supports placement, while cerebellum recovery preserves useful slider motion through short monotonic strokes from the achieved state.The Brain controls task order; the frozen local controller handles interaction recovery, and command completion alone proves neither placement nor motion.Source: §E.6 · pp. 31–32

E.7 SKILL INVOCATION AND THE PVG EVIDENCE BOUNDARY

  • One Proposer uses frozen brain and cerebellum memories to choose task objectives, generate local API programs, inspect evidence, and refresh public bindings without replaying completed actions.The walkthrough is explanatory, not a measured trajectory; unresolved bindings require more public evidence before dependent actions.Source: §E.7 · p. 32
  • Training separates rollouts from persistent updates: the Verifier supplies filtered feedback, the Proposer authors candidates, and the Governor admits updates only after validation.Stage I updates cerebellum memory; Stage II updates brain memory while the cerebellum remains fixed. The retained skill files do not provide a complete candidate-by-candidate PVG transcript or isolate individual rules’ causal effects.Source: §E.7 · p. 33

F PVG ROLE PROMPT TEMPLATES

  • The six edited templates instantiate the three Algorithm 1 roles; they illustrate current PVG responsibilities rather than verbatim historical prompts.The Proposer invokes skills, the Verifier diagnoses outcomes, and the Governor admits validated candidates; API signatures and returns are defined elsewhere.Source: F PVG ROLE PROMPT TEMPLATES · p. 33

F.1 PROPOSER: CEREBELLUM ROLLOUT

  • During Stage I, the Proposer grounds local actions in current public observations, checks returned results and physical effects, and keeps exploratory successes pending until validation supports admission.It records confirmed effects, uncertainty, and API failures rather than treating an issued command as proof of success.Source: §F.1 · pp. 33–34

F.2 PROPOSER: BRAIN ROLLOUT

  • During Stage II, the Proposer composes frozen cerebellum skills to meet task objectives, revises plans from public evidence, and leaves both libraries unchanged during rollout.The cerebellum remains fixed through recovery, candidate construction, and validation of proposed brain updates.Source: §F.2 · p. 35

F.3 PROPOSER: SKILL CONSOLIDATION

  • In a fresh context, the Proposer turns filtered feedback into a candidate with public applicability conditions, hierarchy-appropriate guidance, executable API templates, expected effects, and limitations.The candidate awaits validation; the Proposer neither changes active memory nor claims an admission decision.Source: §F.3 · p. 36

F.4 VERIFIER: DIAGNOSIS AND PUBLIC FEEDBACK

  • The Verifier uses privileged simulation evidence to locate the first unsupported transition, then gives the Proposer public feedback and a focused diagnostic test without revealing hidden state.It diagnoses API, local-manipulation, or task-composition failures, but does not author updates or decide admission.Source: §F.4 · p. 37

F.5 GOVERNOR: VALIDATION AND ADMISSION

  • The Governor checks candidates against public API semantics and subsequent validation, accepting only supported improvements while keeping privileged evidence out of deployed skills.Stage I admission changes only cerebellum memory; Stage II changes only brain memory with cerebellum frozen, and both memories remain fixed during evaluation and deployment.Source: §F.5 · p. 38

F.6 PROPOSER: FROZEN EVALUATION AND DEPLOYMENT

  • During held-out evaluation and deployment, the Proposer uses frozen memories through the shared interface, binds actions from fresh public evidence, and updates only episode working state.The Verifier and Governor are absent, with no online skill-memory updates or task-policy fine-tuning.Source: §F.6 · p. 39

G PVG TRAINING PSEUDOCODE

Algorithm 1 applies a shared PVG training loop to sequential skill acquisition, training a Cerebellum memory before a Brain memory.

G PVG TRAINING PSEUDOCODE

  • The Proposer rolls out on sampled tasks using the simulator, API documentation, and the current skill memory before privileged execution evidence is passed to the Verifier.The Verifier’s feedback is required to belong to the public-feedback set.Source: G PVG TRAINING PSEUDOCODE · p. 40
  • Candidate memory updates are applied only after the Governor accepts them, and training repeats until a prespecified stage-level validation criterion is met.The Governor bases its decision on validation evidence collected for the candidate update.Source: G PVG TRAINING PSEUDOCODE · p. 40
  • After feedback, the Proposer consolidates a candidate update, which is evaluated on subsequent rollouts before the Governor decides whether to accept it.The pseudocode names the validation evidence eδ and the Governor’s accept-or-reject decision.Source: G PVG TRAINING PSEUDOCODE · p. 40
  • The procedure trains the Cerebellum stage first, then trains the Brain stage using the learned Cerebellum memory as its initial context.The algorithm specifies separate task distributions for the two stages.Source: G PVG TRAINING PSEUDOCODE · p. 40
Loading 2610.02788v1…