PaperlayerView exact revision on arXiv

Source-linked AI summary

Playful Agentic Robot Learning

Junyi Zhang, Jiaxin Ge, Hanjun Yoo, Letian Fu, Zihan Yang, Yaowei Liu, Raj Saravanan, Shaofeng Yin, Justin Yu, Dantong Niu, Zirui Wang, Roei Herzig, Ken Goldberg, Yutong Bai, David M. Chan, Ion Stoica, Angjoo Kanazawa, Jiahui Lei, Haiwen Feng, Trevor Darrell

arXiv:2606.19419v1cs.ROcs.AI

TL;DR

Agentic robot systems typically acquire reusable skills only after explicit instructions, limiting continual learning before deployment. RATS uses self-directed play to build a persistent code-skill library, improving downstream performance across benchmarks and enabling transfer to other Code-as-Policy agents.

  • Problem

    Current agentic robot systems remain largely task-driven, acquiring reusable skills only after external instructions rather than through pre-deployment continual learning.

    Source: §1 · p. 2

  • Method

    RATS uses self-directed play to propose, execute, verify, diagnose, retry, and distill exploratory Code-as-Policy behaviors into a persistent skill library.

    Source: §1 · p. 2, §3.1 · p. 4

  • Results

    20.6 and 17.0 percentage-point gains over CaP-Agent0 on LIBERO-PRO and MolmoSpaces show that play-learned skills improve held-out downstream tasks.

    Source: §1 · p. 2, §4.2 · p. 7

  • Takeaways & Limitations

    Play-learned code skills improve agentic robot systems across downstream tasks and can transfer as a plug-and-play addition to other inference-time Code-as-Policy methods.

    Source: §1 · p. 2, §5 · p. 9

    RATS remains primarily simulation-based, and its skill acquisition is constrained by simulation diversity, improper reuse, inference costs, verification dependence, and primitive-level control APIs.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Current agentic robot systems can write executable Code-as-Policy programs, observe feedback, and revise behavior across multiple attempts, but they remain largely task-driven: reusable skills are acquired only after explicit instructions. We study Playful Agentic Robot Learning, where an embodied coding agent uses self-directed play as a continual skill-learning stage before downstream tasks arrive. We introduce RATs, Robotics Agent Teams designed for play-time skill acquisition. During play, RATs proposes novel yet learnable exploratory tasks, plans and executes robot-code policies, verifies intermediate progress, diagnoses failures, retries with dense, step-level feedback, and distills successful executions into a persistent code skill library. At test time, the agent reuses relevant skills from this frozen library to help solve new tasks. Experiments in LIBERO-PRO and MolmoSpaces show that play-learned skills improve held-out downstream tasks over no-play and random-play baselines, with 20.6 and 17.0 percentage-point gains over CaP-Agent0 on LIBERO-PRO and MolmoSpaces, respectively. Moreover, the learned skills can be plugged into other inference-time Code-as-Policy agents by simply retrieving them into the context, improving RoboSuite and real-world transfer by 8.9 and 8.8 points, respectively, without finetuning the underlying model.

1 Introduction

This section introduces Playful Agentic Robot Learning, in which embodied coding agents acquire reusable skills through self-directed play before downstream tasks, and presents RATS as a structured system for proposing, executing, verifying, retrying, diagnosing, and distilling exploratory behaviors into a persistent code library. RATS improves held-out performance and transfers across environments without finetuning the underlying model.

Motivation

  • Playful Agentic Robot Learning uses self-directed play as a continual skill-learning stage before downstream tasks, addressing agentic systems that otherwise learn only after external instructions.The approach is motivated by children’s play and intrinsic-motivation principles favoring experiences that are novel yet learnable.Source: §1 · p. 2

Method

  • RATS proposes exploratory tasks, plans and executes Code-as-Policy programs, verifies progress, diagnoses failures, retries with feedback, and distills successful executions into a persistent code skill library.The library is reused at inference time to solve novel tasks.Source: §1 · p. 2
  • Per-step verification, multi-attempt retry, failure diagnosis, and memory updates provide dense feedback that preserves working policy parts, localizes missing capabilities, and supports reusable skill extraction.Task proposal follows a novelty–learnability rule that targets interactions expanding the skill library while remaining physically achievable [13].Source: §1 · p. 2

Results

  • RATS improves average success by 20.6 percentage points over CaP-Agent0 on LIBERO-PRO and by 17.0 points on MolmoSpaces, while its skills improve RoboSuite success by 8.9 points when plugged into CaP-Agent0.The evaluation tests simulated play environments and held-out downstream benchmark tasks, including cross-environment skill transfer.Source: §1 · p. 2

Contribution

  • The learned code skills function as a plug-and-play addition to other inference-time Code-as-Policy agents, improving downstream performance without finetuning the underlying model.This extends agentic robot learning from reactive task completion toward proactive accumulation of reusable capabilities.Source: §1 · p. 2

2 Related Work

Prior work connects play and curiosity with developmental skill acquisition, continual learning with reusable skills and curricula, and Code-as-Policy with executable, feedback-driven robot programs. RATS unifies these threads by autonomously discovering practice curricula and distilling physically grounded code skills before downstream tasks.

Play, Curiosity, and Developmental Robotics

  • Play and curiosity support self-directed, task-agnostic skill acquisition, which RATS extends to robot programs that execute exploratory goals and distill reusable code skills.Prior approaches pursue novelty, information gain, learning progress, or intermediate difficulty, while robot play learns reusable manipulation behaviors without segmented demonstrations.Source: §2 · p. 3

Continual Skill Learning and Curriculum in Robotics

  • Continual learning and curriculum methods reuse skills and order tasks near competence boundaries, but usually assume predefined tasks, goals, rewards, or experience; RATS autonomously discovers the curriculum.These methods include options, motor primitives, affordances, behavioral priors, hierarchical policies, and goal-conditioned learning over diverse goals.Source: §2 · p. 3

Code-as-Policy and Agentic Robot Learning

  • Code-as-Policy agents synthesize modular robot programs and improve them through feedback, yet typically accumulate reusable skills only while solving given tasks; RATS asks what to practice beforehand.Related systems use pretrained skills, affordances, visual feedback, reasoning-action loops, self-refinement, tool use, and verifiable rewards, while Voyager and sleep-time compute motivate pre-query skill acquisition.Source: §2 · p. 3

3 Robotics Agent Teams for Playful Agentic Robot Learning

RATS enables playful agentic robot learning by autonomously proposing and practicing tasks, executing and refining robot-code policies, and distilling successful behaviors into a reusable skill library. At test time, the frozen library supports standalone Code-as-Policy agents or the full RATS execution team.

3.1 Problem Setup and Overview

  • RATS replaces externally instructed tasks with self-generated play in which reusable skills are acquired through environment interaction and retained for downstream solving.It maintains an initial-plus-learned skill library and failure memory, coordinated by Task Proposer, Execution, and Memory-Management teams.Source: §3.1 · p. 4

3.2 Task Proposer Team

  • The Task Proposer selects novel, learnable tasks using intrinsic curiosity and a Goldilocks objective that favors semi-familiar tasks near a 0.5 success rate.Candidate tasks are generated from scene context, skills, and failure history, then instantiated by an Environment Creator for LIBERO-PRO.Source: §3.2 · pp. 4–5

3.3 Execution Team

  • The Execution Team runs a Write-Execute-Verify-Diagnose loop, using plan checks, code screening, goal and per-step verification, diagnosis, retries, and isolated sub-action practice.Persistent local bottlenecks can trigger a SubAgent to practice a sub-action and return a reusable helper routine.Source: §3.3 · p. 5

3.4 Memory-Management Team

  • The Memory-Management Team extracts successful functions into the skill library while recording failure lessons, tracking reliability, and curating skills and memories over time.Skills are promoted after repeated success, deprecated after repeated failure, and supplemented with experimentally proposed helpers for recurring missing capabilities.Source: §3.4 · pp. 5–6

3.5 Evaluation With the Learned Skill Library

  • After play, intrinsic exploration and memory updates are disabled, while the frozen skill library is reused either by CaP-Agent0 or by RATS for downstream tasks.RATS uses the learned skills to bypass low-level physical bottlenecks and compose plans for complex tasks.Source: §3.5 · p. 6

4 Experimental Results

RATS’s play-learned skills improve in-domain Code-as-Policy performance, transfer across simulation environments and embodiments, and improve preliminary real-world manipulation without finetuning. Curiosity-driven play and the full test-time execution system contribute complementary gains.

4.1 Experiment Details

  • Experiments compare RATS with VLA policies and CaP-Agent0, using task success rate in plug-in and full-system evaluation modes across LIBERO-PRO, MolmoSpaces, and held-out RoboSuite.RATS performs 50 play iterations in each in-domain environment, while RoboSuite is reserved for cross-environment transfer.Source: §4.1 · pp. 6–7, §4 · p. 6

4.2 In-Domain Evaluation

  • RATS raises LIBERO-PRO success from 23.2% to 43.8% (+20.6 pp) and MolmoSpaces success from 21.0% to 38.0%, outperforming baselines and improving across task categories.LIBERO-PRO evaluation uses 600 trials, while MolmoSpaces uses 400 trials; LIBERO-PRO gains extend beyond practiced tasks.Source: §4.2 · p. 7

4.3 Cross-Environment Transfer

  • Transferred LIBERO-PRO skills raise RoboSuite success from 40.3% to 49.1% (+8.9 pp), including gains on cube lifting and two-arm lifting.The skills transfer across environments and embodiments despite being learned on a single-arm robot; RoboSuite was unseen during play.Source: §4.3 · p. 8

4.4 Ablating Play Strategy and Test-Time Execution

  • Curious Play and improved test-time execution are complementary: combining both reaches 44.3%, versus 23.2% with No Play and baseline execution.Under CaP-Agent0, Random Play reaches 24.7% while Curious Play reaches 32.3%; improving only execution reaches 36.3%.Source: §4.4 · p. 8, Table 4 · p. 8

4.5 Sim-to-Real Evaluation

  • RATS exports the LIBERO-PRO skill library directly to a physical robot, improving the CaP-Agent0 baseline by 8.8 percentage points without real-world play or finetuning.The evaluation covers simple manipulation instructions such as cube picking and placing.Source: §4.5 · p. 8, Table 3 · p. 8

5 Conclusion

RATS acquires reusable Code-as-Policy skills through self-proposed play before downstream tasks, improving performance across benchmarks and transferring beyond the original play environment without finetuning the underlying model.

5 Conclusion

  • RATS substantially improves over CaP-Agent0 and VLA baselines on LIBERO-PRO and MolmoSpaces, while random play under the same budget yields much smaller gains.These ablations support the value of self-directed play rather than merely additional interaction.Source: §5 · p. 9
  • Skills learned through self-proposed play transfer across environments, with benefits observed on RoboSuite and in real-world evaluations.The cross-environment results suggest that the acquired skills are reusable beyond the original play environment.Source: §5 · p. 9
  • RATS provides a practical route to improving agentic robot systems without finetuning the underlying model.Its framework acquires reusable Code-as-Policy skills before downstream tasks are provided.Source: §5 · p. 9

Limitations · Appendix · A Implementation Details about RATS

RATS remains limited by simulation-dominated evaluation, constrained practice diversity, mismatched skill reuse, and costly, primitive-level control. Its implementation comprises Task Proposer, Execution, and Memory-Management teams, with learned skills used during evaluation through defined agent prompt interfaces.

Limitations

  • RATS requires larger-scale physical deployment to validate robust sim-to-real transfer because evaluation remains primarily simulation-based.Source: Limitations · p. 9
  • Limited simulation diversity restricts the objects, dynamics, and affordances available for play-time skill acquisition.Source: Limitations · p. 9
  • Improper skill reuse can reduce downstream performance when retrieved skills do not fit the task, motivating better retrieval and context-aware selection.Source: Limitations · p. 9
  • RATS increases inference costs and relies heavily on VLM verification, motivating lighter feedback mechanisms for real-world scaling.Source: Limitations · p. 9
  • Primitive-level control APIs limit dexterous manipulation and motivate richer low-level controllers for real-world scaling.Source: Limitations · p. 9

A Implementation Details about RATS

A.1 Details of Play-Time Task Proposal

RATs proposes executable play tasks from scene context, skill-library summaries, and recent attempts, then selects candidates by novelty and competence while filtering failures and inaccessible objects. Selected proposals are instantiated and deterministically verified before execution, with rejected tasks returned for reconsideration.

Candidate task generation

  • The Task Proposer generates candidate tasks with required objects and skills from scene context, a compact skill-library summary, and recent attempt history.The summary includes skill names, descriptions, reliability tiers, and empirical success statistics without full source code.Source: §A.1 · p. 15
  • Environment-specific grounding limits proposal targets to accessible objects, using manipulation primitives and inventories in LIBERO and visible, reachable objects in MolmoSpaces.MolmoSpaces first converts object identifiers into readable phrases from the current visual observation.Source: §A.1 · p. 15

Goldilocks-driven task selection

  • Candidates are ranked by object-skill novelty times a competence-frontier score, using Wilson lower bounds, while recent failures are downweighted and diagnosable failures enter a bounded retry bank.Retry variants receive a decaying bonus, and initially plausible plans can motivate simplified follow-up tasks.Source: §A.1 · p. 15

Environment creation

  • The Environment Creator converts selected proposals into executable instances, validating LIBERO BDDL syntax and semantics and performing one bounded repair when validation fails.Static checks ensure referenced objects, fixtures, regions, predicates, and assets match the proposal before MuJoCo instantiation.Source: §A.1 · p. 15

Environment verification

  • Before execution, deterministic verification checks successful instantiation, rendering, simulator access, object realization, goal evaluation, and severe-penetration avoidance; rejected tasks return to proposal.MolmoSpaces additionally verifies visual grounding after environment rebinding, preventing rejected tasks from consuming execution attempts.Source: §A.1 · p. 15

A.1.1 Example Trace of Play-Time Task Proposal … B Additional Real-World Experimental Results

RATS combines play-time task proposal, verified code execution, persistent skill and failure-memory updates, and frozen-library evaluation through both full RATS and plug-and-play Code-as-Policy settings. Its task proposer uses prior skills, recent outcomes, and competence-frontier scoring, while execution components localize failures and promote validated reusable helpers.

A.1.1 Example Trace of Play-Time Task Proposal

  • The task proposer combines primitives, learned skills, recent outcomes, and analytical competence-frontier scoring to select exploratory objectives, such as a tissue-box lift that exposed a grasp-centering bottleneck.The proposer receives 13 primitives and nine learned skills, varies from recent attempts, and uses Wilson-bounded reliability in ranking.Source: §A.1.1 · pp. 16–18, Table 7 · p. 18

A.2 Details of RATS for Execution

  • RATS executes accepted tasks through bounded planning, code generation, static checking, goal verification, per-step diagnosis, and targeted retries or SubAgent practice.Failures are routed to plan refinement, local code edits, or physical sub-action practice, while reusable SubAgent routines require successful validation before library promotion.Source: §A.2 · pp. 18–19

A.3 Details of Skill Library and Failure Memory Updates

  • The skill library stores validated executable helpers with metadata, dependencies, provenance, usage statistics, and reliability tiers, prioritizing verified skills and hiding deprecated ones by default.Skills become verified after at least three uses with success rate at least 0.5 and deprecated after at least ten uses with success rate at most 0.2.Source: §A.3 · p. 19
  • Successful trajectories yield parameterized reusable helpers, while failures become structured episodes and compact condition-correction lessons retrieved by task and object overlap.A curator removes redundancy, and the anticipatory Skill Proposer drafts experimental helpers when repeated failures reveal missing capabilities.Source: §A.3 · p. 19

A.4 Details of Evaluation With the Learned Skill Library

  • At test time, the learned library is frozen, task proposal and curation are disabled, and evaluation compares plug-and-play standard Code-as-Policy execution with full RATS execution.The two settings isolate transfer from the learned code library versus the combined value of play-time acquisition and the RATS execution loop.Source: §A.4 · pp. 19–20
  • Plug-and-play CAP-AGENT0 receives retrieved skill signatures, descriptions, source code, and executable definitions, without RATS planning, verification-driven retries, or failure-memory retrieval.A lightweight selector can retrieve a task-relevant subset for large libraries, isolating the transfer value of learned code.Source: §A.4 · p. 20
  • Full RATS conditions planning and policy writing on selected frozen skills while retaining executable dependencies, goal and per-step verification, diagnosis, and bounded retries for unseen tasks.Verified skills are prioritized over experimental ones during plan-step selection.Source: §A.4 · p. 20

B.1 Additional Quantitative Results for Real-World Experiments

MolmoSpaces play-learned skills transfer to additional real-world manipulation tasks when reused with CAP-AGENT0, improving success on object rearrangement and articulated-object interaction.

Task-level results

  • On Swap Cubes, CAP-AGENT0 fails all trials, whereas the skill-augmented agent succeeds in 7 of 30 trials; Close Drawer success increases from 2/30 to 8/30.Swap Cubes requires object rearrangement, while Close Drawer requires articulated-object interaction.Source: §B.1 · p. 20

Additional real-world evaluation

  • +21.7 percentage points: MolmoSpaces skills raise average real-world success from 3.3% to 25.0% across Swap Cubes and Close Drawer.The skills are learned by RATS during MolmoSpaces play and reused with CAP-AGENT0.Source: Table 8 · p. 20, §B.1 · p. 20

B.2 Additional Qualitative Results · C Details about Play Process

Additional real-world rollouts show that CAP-AGENT0 augmented with MolmoSpaces play-learned RATs skills solves object-rearrangement and articulated-object tasks. The supplementary materials also report play-objective statistics and provide videos of real-world and simulation executions.

B.2 Additional Qualitative Results

  • CAP-AGENT0 augmented with MolmoSpaces play-learned RATs skills successfully solves newly added Swap Cubes, Close Drawer, and Open Drawer tasks in the real world.These qualitative results are shown in Figure 4.Source: §B.2 · p. 20
  • In Swap Cubes, the robot exchanges a cube on the platform with a cube below the platform.Source: §B.2 · p. 20
  • In Close Drawer, the robot closes an initially open drawer, while in Open Drawer, it opens a closed drawer.Source: §B.2 · p. 20
  • The MolmoSpaces skill library supports both object rearrangement and articulated-object manipulation in the real world.Source: §B.2 · p. 20
  • Video results include these real-world rollouts and additional simulation rollouts on the project webpage.Source: §B.2 · p. 20
  • Figure 6 summarizes the distribution of play objectives using 49 saved proposal records from the 50-iteration run.Source: Figure 6 · p. 21

C.1 Play-Time Objectives and Knowledge Accumulation

MolmoSpaces play generates diverse, learnable interaction objectives and accumulates reusable skills alongside failure-derived supervision. Across 50 iterations, the library and failure memory expand while successful helpers are reused across related objects and tasks.

Distribution of proposed objectives

  • The proposer covers seven interaction families and more than 30 object categories, with pull and lift dominant alongside articulation, pushing, placing, and picking objectives.This diversity exposes the robot to related but non-identical perception, grasping, motion, and contact problems.Source: §C.1 · p. 21

Practicing novel yet learnable objectives

  • At iteration 25, sliding a black shoe achieves novelty 1.0 and frontier score 0.6756, succeeds after two attempts, and produces a reusable helper.Five iterations later, slide_grasped_object_and_release is reused to slide a Bluetooth speaker, demonstrating practice near the current skill frontier.Source: §C.1 · p. 21
  • A black-metal-rail push helper transfers to a handheld tool in one attempt, while a drawer-handle helper transfers to a drawer pull in one attempt.The transfers occur at iterations 35 and 46 after helpers are learned at iterations 28 and 43, respectively.Source: §C.1 · p. 21

Skill and memory growth

  • From iterations 10 to 50, the skill library grows from 6 to 27 helpers, while failure memory expands from 14 raw episodes and 8 lessons to 70 episodes and 121 lessons.The library tracks experimental, verified, and deprecated helpers, retaining negative evidence while filtering capabilities that fail to generalize.Source: §C.1 · p. 21, Figure 7 · p. 22

Learning from unsuccessful play

  • Failed play produces experimental hypotheses for pull-direction estimation, grasp selection, centroid adjustment, top-down grasps, and approach-axis filtering.Failure memory and the Skill Proposer convert recurring physical bottlenecks into explicit helpers for future practice rather than discarding failed trajectories.Source: §C.1 · p. 22

C.2 Learned-Skill Usage in MolmoSpaces Evaluation · C.3 Qualitative Comparison with Direct Code Synthesis

In MolmoSpaces evaluation, play-learned skills are invoked broadly and composed according to task demands, with successful play-to-evaluation transfer illustrated through cabinet opening. Qualitative comparisons show that RATS succeeds where direct code synthesis fails by composing reusable abstractions instead of reconstructing low-level procedures.

C.2 Learned-Skill Usage in MolmoSpaces Evaluation

  • 391 of 400 evaluation trials invoke learned skills, using 14 of 27 library skills across 5,169 runtime calls.Opening averages 19.0 helper calls per trial, while pick-and-place averages 14.5 and uses broader compositional chains.Source: §C.2 · p. 22, Table 9 · p. 23
  • Learned-skill composition varies by task: localization dominates pick and pick-and-place, while opening and closing combine it with geometry, grasping, and push helpers.Localization accounts for 75.2% of pick calls and 72.3% of pick-and-place calls; direction-and-geometry accounts for 32.3% of opening calls, and push/slide/place helpers account for 21.6% of closing calls.Source: §C.2 · p. 22, Figure 8 · p. 23
  • Object-localization helpers are reused most often, with 2,806 invocations, including 1,873 calls to localize_and_verify_object_point_cloud.The category breakdown makes task-dependent skill compositions explicit rather than limiting the library to perception.Source: §C.2 · p. 22
  • A failed play attempt produced pull-direction and grasp-selection helpers that later enabled a successful drawer-opening iteration and evaluation trial.The helpers get_axis_aligned_pull_direction and select_grasp_for_pulling compose localization, grasp planning, handle grasping, and pulling.Source: §C.2 · pp. 22–23, Figure 9 · p. 24

C.3 Qualitative Comparison with Direct Code Synthesis

  • In a matched drawer-opening comparison, the learned-library run leaves the drawer open, whereas direct synthesis fails to complete the articulation.The learned path composes localization, pull-grasp planning, and handle-motion helpers instead of reconstructing grasp selection, approach geometry, and pull direction.Source: §C.3 · p. 24, Figure 10 · p. 24
  • Across three matched MolmoSpaces objectives, direct CAP-AGENT0 synthesis fails while RATS succeeds by composing learned skills.The objectives are picking up a vintage yellow track shoe, closing the table, and placing a spanner in or on the white bowl blue.Source: §C.3 · p. 25
  • The comparisons show RATS replacing repeated low-level perception, grasp, geometry, and waypoint reconstruction with named reusable control abstractions.Examples include execute_top_down_grasp_and_lift, push_object_closed, verified target localization, and translate_grasped_object_and_release.Source: §C.3 · p. 25

C.4 Usage of LIBERO-Derived Skills in RoboSuite Evaluation · D Details about the Evaluation Benchmark

RATS transfers LIBERO play-time skills to RoboSuite by implicitly reusing them in generated code, enabling completion where direct synthesis fails. Evaluation uses a frozen library on externally specified tasks across two simulated domains.

C.4 Usage of LIBERO-Derived Skills in RoboSuite Evaluation

  • Without the learned library, direct synthesis samples Contact-GraspNet candidates for each handle and scores them with axis and direction heuristics.Source: §C.4 · p. 25
  • The direct-synthesis policy reaches the lift stage but does not complete the RoboSuite objective.Source: §C.4 · p. 25
  • The learned library enables the RoboSuite two_arm_lift policy to complete the task, unlike direct synthesis without the library.Figure 12 compares archived videos and generated code for the same objective.Source: §C.4 · p. 25
  • With RATS, the policy writer selects relevant skills from a frozen library learned from LIBERO play-time tasks.Source: §C.4 · p. 25
  • The transferred skills are reused implicitly rather than invoked as explicit API calls in the final RoboSuite policy.Source: §C.4 · p. 25

D Details about the Evaluation Benchmark

D.1 MolmoSpaces Evaluation Benchmark · D.2 LIBERO-PRO Evaluation Benchmark · E Additional Studies

The evaluations define controlled held-out benchmarks for MolmoSpaces and LIBERO-PRO, with standardized task suites, trial budgets, and success criteria. Qualitatively, RATS succeeds on matched tasks where direct code synthesis fails by reusing verified learned skills.

D.1 MolmoSpaces Evaluation Benchmark

  • MolmoSpaces uses 40 randomly selected held-out episodes spanning opening, closing, picking, and pick-and-place across Open-v1, Close-v1, Pick-v2-classic, and PnP-v2.The main sweep repeats each selected task ten times, producing 400 rollouts.Source: §D.1 · pp. 25–26, Table 10 · p. 27
  • MolmoSpaces episodes span 37 houses and evaluate success with simulator-specific predicates for articulated states, object states, and receptacle-supported placement.Each episode records scene and house metadata, robot initialization, object poses, task state, referral expressions, and camera specifications.Source: §D.1 · p. 25
  • RATS succeeds while CaP-Agent0 fails on matched MolmoSpaces shoe-picking, table-closing, and pick-and-place tasks by reusing verified learned skills.RATS invokes verified localization, grasping, closing, transport, and release skills, whereas CaP-Agent0 reconstructs perception, geometry, pose, and waypoint logic inline.Source: Figure 11 · p. 26

D.2 LIBERO-PRO Evaluation Benchmark

  • LIBERO-PRO evaluates object, goal, and spatial categories under Pos swap perturbations and Task task-specification perturbations, with ten language-conditioned manipulation tasks per suite.The Pos setting changes object positions, while the Task setting changes the task specification.Source: §D.2 · p. 26
  • RATS transfers LIBERO-derived skills to RoboSuite two-arm lifting, succeeding on an evaluation seed where direct code synthesis fails.The transfer uses skills selected from a LIBERO-derived library.Source: Figure 12 · p. 27
  • LIBERO-PRO comparisons cap evaluation at ten initial-state trials per task, yielding 600 rollouts per setup, while baseline and learned-library conditions share the same budget.Although each task provides 50 initial states, the reported comparison uses 6 × 10 × 10 trials.Source: §D.2 · p. 27

E.1 Ablation Results on MolmoSpaces

On MolmoSpaces, play-learned skills improve performance under both the standard CAP-AGENT0 agent and the full RATS execution system, with larger gains when RATS retrieves, verifies, and composes them.

Ablation Setup

  • All play-based variants use 50 play iterations, with play-time skills learned by the proposed RATS system.The ablation evaluates play strategies and test-time systems on open, close, pick, and pick-and-place MolmoSpaces tasks.Source: §E.1 · p. 27, Table 12 · p. 28

Ablation Results

  • 38.0% with Curious Play versus 32.8% without play under the full RATS execution system, while CAP-AGENT0 improves from 21.0% to 25.8%.The results indicate complementary benefits from play-time skill learning and structured test-time execution.Source: §E.1 · pp. 27–28, Table 12 · p. 28

E.2 Token Cost Analysis … F Agent I/O and Prompts

RATS converts play-time computation into reusable skills that outperform compute-matched extra retries, while its agent interfaces and prompts specify how tasks, environments, and policies are generated. The appendices also expose representative MolmoSpaces and LIBERO skills implementing perception, manipulation, verification, and correction routines.

E.2 Token Cost Analysis

  • Failed iterations exhausting five-attempt retry budgets dominate play-time tokens, with Write-Execute-Verify-Diagnose revisions costing more than intrinsic task proposal.This analysis was conducted in the libero_spatial environment, and Table 13 breaks down the component-level consumption.Source: §E.2 · p. 28, Table 13 · p. 29
  • RATS Skills achieves 32.3% average success, versus 26.0% for compute-matched CAP-AGENT0 with 15 turns and 23.2% for the standard baseline, while retaining the standard 10-turn test-time system.The skills are learned from 50 iterations of Curious Play; the 26.0% baseline receives additional retry budget to match amortized token cost.Source: Table 14 · p. 29, §E.2 · p. 29

E.3 MolmoSpaces Learned Skills

  • MolmoSpaces skills include verified axis-aligned pull-direction computation, experimental object-closing through localization and inward pushing, and a deprecated wrist-camera top-down grasp planner.The examples report 30 uses and 10 successes for the pull-direction skill, 2 uses and 2 successes for closing, and 26 uses and 6 successes for grasp planning.Source: §E.3 · pp. 29–30

E.4 LIBERO Learned Skills

  • LIBERO skills combine pose-alias and segmentation-fallback localization, configurable OBB-based grasp retries with lift verification, and hover-based target re-segmentation for corrected placement.The snippets show verified localization and placement routines, plus an experimental pick routine that records attempts and returns success only after detecting sufficient lift.Source: §E.4 · pp. 31–34

F Agent I/O and Prompts

  • Agent I/O specifications enumerate the inputs and outputs of RATS’s agents, while prompts instantiate LIBERO experiments and adapt to MolmoSpaces through environment-specific APIs, observations, and scene grounding.Runtime-specific scene descriptions, API documentation, retrieved memories, visual evidence, and task code are represented by placeholders in the prompt documentation.Source: Table 15 · p. 35, F Agent I/O and Prompts · p. 35, §F.1 · p. 35

F.2 Agent Prompts

  • The Task Proposer prompt directs curiosity-driven, single-step, child-feasible exploration using skill history, known limitations, reliable pick lists, valid predicates, and expected-skill outputs.It requires JSON fields for reasoning, language, scene, objects, fixtures, goals, novelty, difficulty, and affordance hints, while forbidding unsupported Stack predicates.Source: §F.2 · pp. 36–37
  • The Environment Creator prompt converts task proposals into valid MuJoCo BDDL by enforcing catalog scene types, placement regions, prefixed region naming, exact object types, and fixture-specific subregions.The supplied microwave example demonstrates heating_region and top_side requirements, complete initial placements, and the final In and Close goals.Source: §F.2 · pp. 37–39

Related papers

Loading 2606.19419v1…