Experience-Guided Initiation Search for Learned Skills in Skill Composition

arXiv:2610.11418 · cs.RO · Submitted 2026-10-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Today's paper: "Experience-Guided Initiation Search for Learned Skills in Skill Composition".

Dev: The gist: EVIS, an Experience-Guided and BehaviorValidated Initiation Search framework,

Rosa: First, who's behind it and why it matters.

Title and authors: Rosa: So we’re looking at this paper today called "Experience-Guided Initiation Search for Learned Skills in Skill Composition". It’s by Qixuan Li, Yanhong Zhao, and Jincheng Yu. Basically, they're tackling the problem of using frozen learned skills in new places. The core issue is that just because a sequence of skills looks valid symbolically doesn't mean it will actually work when you move to a different environment.

Dev: Yeah, and the authors are proposing this framework called EVIS to find those reliable initiation configurations without needing tons of costly target interactions. They’re focusing on how historical execution data can guide the search process first before testing anything in the new world.

Taro: It sounds like they’re trying to bridge that gap between what worked before and what will work now, which is a big challenge when dealing with complex physical states passed between skills, you know?

Rosa: Exactly. The paper suggests that instead of just looking at historical data as a final answer, you use it to propose candidates for initiation configurations and then you test those proposals in the new environment to see if they actually hold up.

Dev: Right, so the main idea is separating the search strategy from the validation step; history tells us where to look first, and target outcomes tell us if we trust that candidate.

The paper's summary: Rosa: The EVIS framework uses historical execution records to create a set of promising candidates, which they call historical proposals. These proposals are based on context-specific Gaussian process classifiers and a retrieval-weighted predictor, which essentially just ranks the initiation configurations based on what worked previously.

Dev: And then they take those ranked candidates and run them in the target environment to see if they lead to a successful rollout under the objective G. If it succeeds, it gets validated as a provisional candidate, but you don't accept it immediately.

Taro: That sounds smart because instead of throwing every possible configuration at the wall in the new place, you’re narrowing down the search based on past performance first. But what happens when those historical proposals aren't good enough for the new environment?

Rosa: Well, that’s where they have a recovery search stage. If a provisional candidate gets rejected or if there aren't enough good historical proposals, EVIS doesn't just stop; it modifies where the search proceeds to try and find more options.

Dev: They use target-side observations accumulated during the discovery process to redirect the remaining search, which means they can broaden their search geometrically by selecting a configuration that is far away from what they've already found so far.

The paper's improvements: Rosa: One of the key contributions here is treating the initiation configuration as an explicit interface for task-level planning over frozen learned skills, which means it makes finding reliable configurations in new environments much more structured. They aim to do this with as few costly target-environment rollouts as possible.

Dev: They also introduced a finite-budget decision rule, Assessk(z), which acts like a screening mechanism. It tells the system whether to accept a provisional candidate based on some score and whether it should keep searching or reject it based on failure counts and acceptance rates.

Taro: That sounds like they’re building in robustness for when things go wrong; if the initial historical guidance fails, the recovery search is supposed to pull them back into a promising area instead of just giving up.

Rosa: Precisely. They show that combining historical guidance with target-side behavioral validation helps reduce mean target-environment queries significantly. For example, in single-skill discovery, they reduced those queries by fifty-nine point four percent compared to a baseline method using fixed historical reuse while keeping the confirmation outcomes the same.

Dev: And for multi-stage tasks, they reported reliable rates of forty-one point seven percent and fifty-eight point three percent at a budget of B = ten which is much better than the history-agnostic Scratch-Sobol method that only hit an eight point three percent reliable rate.

Conclusion: Rosa: To wrap this up, the paper on Experience-Guided Initiation Search for Learned Skills in Skill Composition shows that we can use historical data to prioritize where to look and target outcomes to validate if those candidates still work when transferring skills. This combination really cuts down on how much interaction is needed in a new environment.

Dev: The main implication is that you don't need perfect transfer; you just need a reliable way to screen past experience against the current environment's behavior, and EVIS does that by using a search policy z t = At(DH, D t-one) based on historical data DH and accumulated target observations D t-one <ref:2610.11418#pg1>.

Taro: From my side, it’s about building a system that can actually handle when the world misbehaves. The recovery search part is important because it ensures that if the initial historical proposals fail, we don't get stuck in a local optimum and we can explore configurations far from what we already found.

Rosa: So, by using this Experience-Guided and BehaviorValidated Initiation Search framework, they’ve shown how historical experience can be reused effectively without having to assume perfect transfer to a new environment. That's the summary of the work on Experience-Guided Initiation Search for Learned Skills in Skill Composition.

cs.RO

Submitted: 2026-10-08

Updated: 2026-10-08

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 80/100

The gist: The gist: EVIS, an Experience-Guided and BehaviorValidated Initiation Search framework, discovers reliable initiation configurations for frozen learned skills in new environments by using historical

Key concepts

Initiation Configuration
This refers to the specific settings or starting conditions required before a learned skill can begin executing. It's crucial because changing this configuration affects not only the current skill but also the physical state passed to all subsequent skills in a sequence, meaning local success doesn't guarantee overall task success.
Experience-Guided Proposal
EVIS uses past execution records to generate a list of potential starting configurations. It ranks these candidates based on historical performance, transferring the ordering information from old data rather than relying on the predictions themselves. This guides the search toward configurations that have been successful in similar contexts.
Behavioral Validation
Once a configuration is proposed by history, it must be tested in a new environment. EVIS uses these target-environment rollouts to validate if the historical candidate is still effective. This step separates historical guidance (where to look) from behavioral evidence (whether to trust the candidate).
Budgeted Return Rule
EVIS employs a specific rule to decide whether a provisional skill configuration should be accepted, even after initial validation. This rule uses a finite budget of target interactions ($K_{max}$) and assesses the trade-off between accepting a promising candidate and continuing the search.

Terminology

Summary

The gist: EVIS, an Experience-Guided and BehaviorValidated Initiation Search framework, discovers reliable initiation configurations for frozen learned skills in new environments by using historical execution experience to prioritize candidates and target-environment rollouts to validate their effectiveness.

Problem Formulation

The core challenge addressed is that a symbolically valid skill sequence may not be physically executable because the initiation configuration of an upstream skill affects both its own execution and the physical state passed to subsequent skills The central question is how historical experience can help discover reliable initiation configurations with fewer target interactions, while remaining robust to imperfect transfer The problem involves finding a configuration z that maximizes the success probability pG(c⋆, z) under a budget of at most Kmax target executions.

EVIS Framework

EVIS treats historical experience as a source of candidate proposals rather than final decisions Historical execution records prioritize promising initiation configurations, while behavioral evidence from target-environment rollouts validates whether these candidates remain effective This separates transfer from decision-making: history determines where to look first, whereas target-side outcomes determine whether a candidate should be trusted.

Experience-Guided Proposal

EVIS constructs historical proposals over a finite candidate set Zcand Historical predictors are evaluated on Zcand to rank candidate configurations, even when the original historical observations were collected at different configurations EVIS transfers historical ordering information rather than treating historical predictions as target-environment decisions. The mechanism uses context-specific Gaussian process classifiers and a retrieval-weighted predictor to construct scoring hypotheses, such as sret(c⋆, z). The proposal mechanism is insensitive to the absolute scale of a historical predictor and depends only on the ordering it induces.

Behavioral Validation and Budgeted Return

When an evaluated initiation configuration z produces a successful target-environment rollout, EVIS treats it as a provisional candidate rather than returning it immediately Additional target-side evidence may be required before the candidate is accepted. EVIS applies the finite-budget decision rule Assessk(z) = accept if Sk ≥ mq, reject if Fk > nq − mq, and continue otherwise, where Kused = Ksearch + Kqual ≤ Kmax. This qualification rule serves as a finite-budget behavioral screening mechanism to determine whether a provisional candidate should be accepted.

Behavior-Guided Recovery Search

If a provisional candidate is rejected or if historical proposals are insufficient, EVIS uses the target-side observations accumulated during discovery to redirect the remaining search The recovery stage modifies where the search proceeds, while leaving the frozen execution policies unchanged. When historical proposals are insufficient, EVIS can broaden the search geometrically using a coverage query to select an untested configuration farthest from the existing discovery set. The regularized target correction estimates a residual function delta t and results in a target-corrected score st(z) = clip[ϵs, 1−ϵs] sH(z) + δt(z).

Experimental Results

EVIS reduces mean target-environment queries and improves reliable candidate discovery under small interaction budgets For single-skill discovery, EVIS reduces target-environment interaction by 59.4% compared to a fixed-budget historical reuse baseline while maintaining identical confirmation outcomes. In multi-stage initiation search, EVIS achieves Reliable Rates of 41.7% and 58.3% at budget B = 10, significantly outperforming the history-agnostic Scratch-Sobol method which achieved only 8.3% Reliable Rate. Mechanism analysis confirms that historical experience provides transferable ranking structure, behavioral validation eliminates all false accepts in both tasks, and recovery search restores coverage after rejected reuse.

The combination of historical guidance with target-side behavioral validation effectively reduces the interaction cost of deploying frozen learned skills in new environments EVIS demonstrates that historical experience can be reused effectively without assuming perfect transfer to a new environment. The results show that combining historical guidance with targetside behavioral validation can reduce the interaction cost of deploying frozen learned skills in new environments.

--- Page 1 ---

Experience-Guided Initiation Search for Learned Skills in Skill Composition Qixuan Li1, Yanhong Zhao2, Jincheng Yu3 Abstract— Deploying frozen learned skills, such as VisionLanguage-Action (VLA) policies, in new environments requires identifying initiation configurations that support reliable execution. In skill composition, an initiation configuration affects not only the current skill but also the physical state passed to subsequent skills, so successful execution of an individual skill does not necessarily imply successful completion of the composed task. Estimating target-specific capability through extensive rollouts is costly in real-world deployment, while directly reusing historical experience can be unreliable under environment changes. We propose EVIS, an Experience-Guided and BehaviorValidated Initiation Search framework for discovering reliable initiation configurations under limited target interaction. EVIS uses historical execution experience to prioritize promising candidates and target-environment behavior to validate whether they remain effective. We evaluate EVIS on single-skill and two-stage manipulation tasks with frozen VLA policies. EVIS reduces mean target-environment queries and improves reliable candidate discovery under small interaction budgets. These results show that combining historical guidance with targetside behavioral validation can reduce the interaction cost of deploying frozen learned skills in new environments.

I. INTRODUCTION Learned visuomotor skills, including pretrained VisionLanguage-Action (VLA) policies, provide reusable manipulation capabilities and are increasingly composed in hierarchical systems for longer-horizon tasks [1]–[4] However, a symbolically valid skill sequence is not necessarily physically executable, because the initiation configuration of an upstream skill affects both its own execution and the physical state handed to subsequent skills. An upstream skill may therefore succeed locally yet leave the system in a state from which the next skill cannot proceed; for example, successfully opening a drawer does not guarantee successful subsequent retrieval or placement Thus, composing frozen learned skills requires determining not only which skill to invoke, but also from which initiation configuration it should be executed.

Determining such initiation configurations in a new environment is challenging. Existing Task and Motion Planning (TAMP) methods primarily describe action applicability through geometric constraints and symbolic preconditions [5], [6] while learned precondition models estimate skill applicability from execution data [7], [8] Recent memorybased methods further reuse historical execution experience to assist the invocation and composition of frozen policies [2], [3] However, for a frozen black-box skill in a new environment, the reliability of different initiation configurations is generally unknown Estimating target-specific capability through extensive rollouts requires substantial interaction, while directly reusing historical candidates may fail as environment geometry, visual appearance, and interaction conditions change Historical experience is therefore useful as guidance, but insufficient as a decision criterion This leads to the central question of this work: how can historical experience help discover reliable initiation configurations with fewer target interactions, while remaining robust to imperfect transfer?

To address this problem, we propose EVIS, an ExperienceGuided and BehaviorValidated Initiation Search framework. EVIS treats historical experience as a source of candidate proposals rather than final decisions Historical execution records prioritize promising initiation configurations, while behavioral evidence from target-environment rollouts—i.e., observed outcomes under the intended execution objective—validates whether these candidates remain effective. This separates transfer from decision-making: history determines where to look first, whereas target-side outcomes determine whether a candidate should be trusted.

II. RELATED WORK A. Skill Applicability and Composition Skill applicability and composition have long been studied in hierarchical control and planning. The options framework formalizes skill initiation and termination [11] while skill chaining learns skills that reach the initiation sets of subsequent skills [12] In task and motion planning, PDDLStream integrates symbolic planning with black-box samplers for continuous feasibility constraints [6] Learned precondition models further estimate manipulation applicability from execution data using object relations [8] or point-cloud representations for contact-rich controllers [7] Unlike these approaches, which primarily model skill applicability or feasibility, EVIS searches for effective initiation configurations of frozen learned skills under limited target interaction, evaluating them by whether they support successful execution of the composed skill sequence.

B. Policy Orchestration and Execution Memory Language-guided robotic systems connect high-level task reasoning with low-level execution in different ways. SayCan grounds skill selection in learned affordances [13] Inner Monologue incorporates online execution feedback into highlevel planning [4] and Code as Policies composes perception and control interfaces through generated programs [14] More recent execution-memory systems reuse past experience to organize and invoke frozen VLA primitives HarnessVLA reuses task-specific execution traces and persistent execution memory to guide primitive composition across deployment variations [3] while RoboHarness uses execution memory and online evidence for capability-aware policy routing and compatible skill handoffs [2] EVIS addresses a complementary problem: it transfers execution experience across environments specifically to guide the search over initiation configurations, while using a small number of target-environment rollouts to validate whether transferred candidates remain reliable.

C. Experience-Guided Search and Candidate Discovery Historical information can accelerate expensive blackbox search.

Improvements for AI systems

  1. This system can reduce mean target-environment queries by using historical experience to prioritize promising initiation configurations, as EVIS uses historical execution experience to prioritize promising candidates.

  2. The improved system can improve reliable candidate discovery under tight budgets by employing a two-stage validation process: Historical guidance improves early candidate discovery under small interaction budgets. Behavioral validation filters unreliable transferred candidates.

  3. The system can achieve higher reliability in task completion by incorporating target-side feedback, as the framework determines whether a candidate should be trusted based on observed outcomes under the intended execution objective.

  4. The architecture supports improved robustness to environment changes through two complementary mechanisms: Experience-guided proposal uses previous executions to prioritize where to search first and behavioral evidence from target-environment rollouts—i.e., observed outcomes under the intended execution objective—validates whether these candidates remain effective.

  5. The system can recover lost coverage when initial historical proposals fail by implementing a recovery search that utilizes Hypothesis disagreement and Coverage and local exploration, allowing it to select configurations based on where historical hypotheses disagree or to select the configuration farthest from the existing discovery set.

  6. The improved system can refine historical proposals using a mechanism that addresses transfer uncertainty by estimating a residual function: it defines the target-corrected score as st(z) = clip[ϵs, 1−ϵs] sH(z) + δt(z), where "δt" estimates the deviation from the transferred prediction.

Abstract

Deploying frozen learned skills, such as Vision-Language-Action (VLA) policies, in new environments requires identifying initiation configurations that support reliable execution. In skill composition, an initiation configuration affects not only the current skill but also the physical state passed to subsequent skills, so successful execution of an individual skill does not necessarily imply successful completion of the composed task. Estimating target-specific capability through extensive rollouts is costly in real-world deployment, while directly reusing historical experience can be unreliable under environment changes. We propose EVIS, an Experience-Guided and Behavior-Validated Initiation Search framework for discovering reliable initiation configurations under limited target interaction. EVIS uses historical execution experience to prioritize promising candidates and target-environment behavior to validate whether they remain effective. We evaluate EVIS on single-skill and two-stage manipulation tasks with frozen VLA policies. EVIS reduces mean target-environment queries and improves reliable candidate discovery under small interaction budgets. These results show that combining historical guidance with target-side behavioral validation can reduce the interaction cost of deploying frozen learned skills in new environments.

Sources

Related papers