Explore, Execute, Evolve: A Skill Acquisition and Reuse Loop for Embodied Agents

1 Introduction

• General-purpose multimodal agents show zero-shot robotic task-solving abilities, but executions often restart from scratch, requiring repeated perception, physical trial and error, and recovery.VLA and WAM models also face limitations in generalizing to unseen tasks.Source: §1 · p. 1
• Physical skill reuse must account for partial observations, changing scenes, embodiment constraints, and uncertain action outcomes, so agents must gather information, verify effects, and recover from deviations.Unlike software interactions, executing a physical command does not guarantee its intended effect [7].Source: §1 · pp. 1–2
• RoboSkill connects exploration, feedback-adaptive execution, and skill updates from execution records; acquired skills then guide later exploration and execution.Outcomes from subsequent cycles inform further skill updates.Source: §1 · p. 2
• Tactile feedback complements vision with evidence of contact and collision, while executable code preserves adaptable procedures alongside textual skill guidance.Text describes strategies and applicability conditions; code retains procedures for related scenes and tasks.Source: §1 · p. 2
• RoboSkill improves task success and execution efficiency in evaluations on LIBERO-10 and real robots, with skills reused across tasks and agents.Further analyses examine tactile feedback, reusable code, and successive skill updates.Source: §1 · p. 2

Vision-Language-Action and World-Action Models

• VLA and WAM models have grown more general, but remain shaped by the tasks and data encountered during training; RoboSkill instead uses deployment-time reasoning and adaptation.WAMs incorporate predictive world modeling to capture physical dynamics and improve action generation [1] [13] [20] [28].Source: §2 · p. 3

Language-Model Agents for Robotics

• Robotic language-model agents range from executable program generation to interactive coding, analytic primitives, data collection, and iterative skill-library exploration.RoboSkill emphasizes transferable text-and-code experience retrievable and adaptable across tasks and execution agents.Source: §2 · p. 3

3 Method

• RoboSkill organizes a shared skill library around exploration, closed-loop execution, and evolution, so skills can guide subsequent exploration.The loop overview is followed by descriptions of exploration, execution, and evolution.Source: §3 · p. 3

3.1 Framework Overview

• The agent receives task instructions, selected skills, and SDK access, then explores for missing information, executes code with feedback, and reviews session records to create or update skills.Updated skills provide guidance for subsequent exploration and execution.Source: §3.1 · p. 3

3.2 Exploration

• Exploration treats needed information as a subtask, using prior findings and skills to guide repeated investigation whose knowledge and procedures become priors for the main task.In the table-height example, contact resolves uncertainty in an RGB-D estimate at approximately 0.446 m before a subsequent grasp attempt.Source: §3.2 · pp. 3–4
• Exploration findings remain in session context for continued execution or are consolidated with records and code into a skill if the agent exits after exploration.Source: §3.2 · p. 4

3.3 Closed-loop Execution

• During execution, the agent adapts selected skill procedures to the scene and uses RGB, depth, and tactile observations to assess action outcomes.Tactile feedback provides contact evidence that may be difficult to determine from images alone.Source: §3.3 · pp. 4–5
• An SDK lets the agent compose reusable executable procedures during interaction, and the resulting code and interaction records feed into skill evolution.Examples include converting depth images to point clouds and verifying table height by descending until contact.Source: §3.3 · p. 5

3.4 Evolution

• Evolution consolidates exploration and task records into reusable skills, with one review at task end to avoid the overhead of reviewing every completed subtask.Skills combine textual strategy and applicability guidance, reusable code, supporting records, and metadata; agents select up to K=3 packages in the experiments.Source: §3.4 · p. 5

4.1 Experimental Setup

• Simulation uses LIBERO-10 with four evaluation seeds per task; real-robot evaluation covers 12 tasks across five stations, with ten trials per task.Baseline uses no saved skills; RoboSkill selects up to three skills in simulation, and reported real-robot average time covers successful trials only.Source: §4.1 · pp. 5–6

4.2 Effectiveness of RoboSkill

• RoboSkill raises simulation Final success for GPT-5.6 Sol (82.5% to 97.5%) and Fable 5.1 (97.5% to 100.0%), and Astra real-robot success on easy and hard tasks (91.7% to 100.0%; 80.0% to 88.3%).Opus SR@30m rises from 17.5% to 82.5%; Astra average time over successful trials falls from 20.3 to 16.6 minutes on easy tasks and 27.0 to 23.1 minutes on hard tasks.Source: Table 1 · p. 6, Table 2 · p. 6, §4.2 · p. 7

4.3 Cross-Task Skill Transfer

• Adding up to three selected skills raises first-episode success for Astra, Fable, and Opus by 5.0, 7.5, and 12.5 points, but Sol falls from 67.5% to 52.5%; selection and skill count co-vary.Source: Table 3 · p. 7, §4.3 · p. 7
• A skill learned on another task raises Sol first-episode success from 30.0% to 44.0% and lowers average time for Sol and Opus, while Opus first-episode success falls from 60.0% to 56.0%.Source: Table 4 · p. 7, §4.3 · p. 8

4.4 Cross-Agent Skill Transfer

• Cross-agent skills raise Kimi K3 first-episode success from 12.0% to 42.0–66.0% in simulation and GPT-5.6 Sol success from 13.3% to 70.0% on real-robot easy tasks.In simulation, Fable skills yield the highest first-episode success, while Opus skills yield the shortest average time.Source: Table 5 · p. 8, §4.4 · p. 8, Table 6 · p. 8

4.5 Skill Evolution

• Skill updates are non-monotonic: Sol’s second update raises first-episode success from 58.0% to 70.0%, while Astra’s real-robot update raises success from 88.3% to 93.3% but increases successful-trial time from 23.1 to 26.2 minutes.Opus reaches its best performance after the first update; a second update maintains 100.0% final success but worsens SR@30m and average time.Source: §4.5 · pp. 8–9, Table 7 · p. 9, Table 8 · p. 9

4.6 Ablation Studies

• Executable code reduces average execution time for all four agents and improves or preserves first-episode success for three; Sol rises from 15.0% to 67.5% success and falls from 67.1 to 39.9 minutes.Source: §4.6 · p. 9, Table 9 · p. 10
• Tactile input cuts simulation average execution time from 25.4 to 15.9 minutes despite first-episode success falling from 82.0% to 78.0%; on the real-robot task, success rises from 80.0% to 90.0%.Real-robot average successful execution time falls from 33.6 to 20.0 minutes on Remove Test Tube.Source: §4.6 · pp. 9–10, Table 10 · p. 10

5 Hardware Setup

• Only the right arm is used in the real-robot experiments.The left arm is visible but unused.Source: §5 · p. 10
• The right arm is equipped with a wrist-mounted camera and a tactile-equipped gripper.Source: §5 · p. 10
• An external overhead camera provides a third-person view of the workspace.Source: §5 · p. 10

6.1 Evaluation Seeds

• The main comparison and same-task ablations construct skills on seed 0 and evaluate on seeds 1–4, yielding 40 task–seed pairs per setting.The construction seed is excluded from both Baseline and skill-assisted conditions.Source: §6.1 · p. 10, Table 11 · p. 11
• Cross-task and cross-agent transfer use all five seeds, 0–4, yielding 50 pairs per setting; skill-refinement experiments instead use seeds 100–104.Cross-task transfer supplies a skill from another task, while cross-agent transfer supplies one from another agent.Source: §6.1 · p. 10

6.2 Simulation Tasks and Evaluation

• LIBERO-10 evaluation covers ten tasks, with each task–seed cell given a four-hour budget and potentially multiple episodes.Episode 1 is the first attempt; each reset starts a new episode, and evaluation ends on success, agent termination, or budget expiration.Source: §6.2 · p. 11, Figure 4 · p. 12
• Final SR, 1st SR, and SR@30m measure success by termination, within the first episode, and within 30 minutes, respectively.Avg. Ep. and Avg. Time are averaged over all cells, including failures.Source: §6.2 · p. 11

6.3 Cross-Task and Cross-Agent Transfer

• Cross-task transfer uses five fixed source–target task pairs in both directions, creating ten directed combinations.Other-Task Skill supplies the assigned source skill without a target-task skill; Baseline has no saved skills.Source: §6.3 · p. 12
• Simulation cross-agent transfer gives Kimi K3 a same-task skill built by one of four source agents; real-robot transfer gives Sol Astra-built skills on six easy tasks.The real-robot transfer uses five trials per task.Source: §6.3 · p. 12

6.4 Real-Robot Tasks and Evaluation

• Real-robot trials use five parallel Piper-arm stations and one episode with a one-hour budget per trial.An on-site volunteer determines success; Astra uses ten trials per task and Sol uses five per easy task.Source: §6.4 · p. 12, Figure 5 · p. 13, Table 12 · p. 13

7.1 Task-Wise Real-Robot Results

• Table 13 reports per-task Astra results before and after exploration for six easy and six hard tasks.The table caption specifies ten trials per task for success rates and one hour per trial.Source: §7.1 · p. 14, Table 13 · p. 14
• Transferred skills enable GPT-5.6 Sol to succeed on all six easy tasks, including three with no recorded Baseline successes.Table 14 covers five trials per task with tactile input and a one-hour trial limit.Source: Table 14 · p. 14, §7.1 · p. 14
• Table 14 reports task-wise GPT-5.6 Sol results before and after receiving skills constructed by Astra.Average time is calculated over successful trials only.Source: Table 14 · p. 14

8.1 Selected Skills

• All four models retain the same-task skill first, while selection can add skills from other tasks or omit rule-selected skills.The agent may select up to three skills; the table lists source-task IDs in selection order.Source: §8.1 · p. 14, Table 15 · p. 15
• All models omit the third rule-selected skill for T4, T6, and T8; mean selected skill counts are 2.6 for the fixed rule and 2.4, 2.3, 2.4, and 1.8 for Sol, Opus, Fable, and Astra.Astra selects only the same-task skill for T3, T5, and T9.Source: §8.1 · p. 15

8.2 Selection Performance and Interpretation

• Agent selection improves first-episode success for Astra, Fable, and Opus but reduces it for Sol; average time falls for Astra and Opus and rises for Sol and Fable.The comparison does not dominate the fixed rule across all models and metrics.Source: Table 16 · p. 15, §8.2 · p. 15
• The comparison cannot isolate skill-count effects from skill relevance, and selection lists do not reveal which skill components agents use.They also do not establish each selected skill’s contribution.Source: §8.2 · p. 15

• RoboSkill reuses skills to guide subsequent exploration and execution, then updates them from new outcomes.Source: §9 · p. 16
• Tactile feedback reduces interaction uncertainty, while reusable code reduces reasoning overhead during skill reuse.Source: §9 · p. 16
• Experiments on LIBERO-10 and real robots demonstrate improved success rates and execution efficiency, with skills supporting reuse across tasks and agents.Source: §9 · p. 16