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

Sicheng Xie; Yitong Chen; Haidong Cao; Shunlin Lu; Zuxuan Wu; Yu-Gang Jiang

cs.ROSep 29, 2026

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

Authors: Sicheng Xie, Yitong Chen, Haidong Cao, Shunlin Lu, Zuxuan Wu, Yu-Gang Jiang

Organizations: Institute of Trustworthy Embodied AI, Fudan University · Shanghai Innovation Institute · NeoteAI.

Abstract

Vision-language-action and world-action models have demonstrated impressive capabilities in robotics, yet generalization to unseen tasks remains challenging. More recently, general-purpose multimodal agents have shown great potential for zero-shot robotic task solving. However, they often incur high execution costs by reasoning and exploring the physical world from scratch. To reduce these costs, we introduce RoboSkill, a framework that connects skill acquisition and reuse through an Explore, Execute, Evolve loop. Within this loop, the agent explores to gather task-relevant information, executes tasks while adapting to feedback, and evolves its skill library based on execution records. It then reuses these skills to guide exploration and execution in the next cycle, closing the loop. To improve loop efficiency, we complement vision with tactile feedback to reduce uncertainty during physical interaction. We further augment textual guidance with reusable code to reduce reasoning overhead during skill reuse. On LIBERO-10, RoboSkill improves first-episode success rates by 12.5--25.0 percentage points and reduces average runtime by 7.6--72.4% across four agents. On real robots, it improves success rates by 8.3 percentage points and reduces average runtime for successful trials by at least 14.4%.

Figures & tables

Figure 1 : RoboSkill accelerates agent execution through skill reuse. In this tilted-nail example, a VLA/WAM follows the vertical hammering strategy learned from its training data. An embodied agent reasons about both grasping and striking, while RoboSkill accelerates execution by reusing the grasping skill and reasoning only about the strike direction.

Figure 2 : The RoboSkill framework. The agent explores to acquire missing information, executes with visual and tactile feedback, and produce skills to guide the next cycle. The example illustrates a nested exploration loop that verifies table height through contact before resuming grasping.

Model / setting	Final SR $\uparrow$	1st SR $\uparrow$	SR@30m $\uparrow$	Avg. Ep. $\downarrow$	Avg. Time (min) $\downarrow$
GPT-6 Astra
Baseline	100.0%	72.5%	87.5%	1.5	18.5
$\mathit{RoboSkill}$	100.0%	97.5%	95.0%	1.0	17.1
GPT-5.6 Sol
Baseline	82.5%	27.5%	32.5%	9.2	94.5
$\mathit{RoboSkill}$	97.5%	52.5%	72.5%	2.9	35.9

Table 1 : Main results on Libero-10.

	Easy Tasks		Hard Tasks
Setting	1st SR $\uparrow$	Avg. Time (min) $\downarrow$	1st SR $\uparrow$	Avg. Time (min) $\downarrow$
Baseline	91.7%	20.3	80.0%	27.0
$\mathit{RoboSkill}$	100.0%	16.6	88.3%	23.1

Table 2 : Main results on real-world using GPT-6 Astra.

	GPT-6 Astra		GPT-5.6 Sol		Fable 5.1		Opus 5
Setting	1st SR $\uparrow$	Avg. Time (min) $\downarrow$	1st SR $\uparrow$	Avg. Time (min) $\downarrow$	1st SR $\uparrow$	Avg. Time (min) $\downarrow$	1st SR $\uparrow$	Avg. Time (min) $\downarrow$
$\mathit{RoboSkill}^{-}$	92.5%	16.5	67.5%	39.9	90.0%	13.0	67.5%	30.4
$\mathit{RoboSkill}$	97.5%	17.1	52.5%	35.9	97.5%	12.5	80.0%	22.3

Table 3 : Using one skill versus multiple skills on LIBERO-10.

\mathit{RoboSkill}^{-}

uses only the skill learned on the target task.

\mathit{RoboSkill}

can additionally select skills learned on other tasks, using up to three skills in total.

	GPT-5.6 Sol		Opus 5
Setting	1st SR $\uparrow$	Avg. Time (min) $\downarrow$	1st SR $\uparrow$	Avg. Time (min) $\downarrow$
Baseline	30.0%	86.0	60.0%	77.0
Other-Task Skill	44.0%	66.3	56.0%	43.5

Table 4 : Can skills transfer to a different task? Other-Task Skill provides one skill learned on a paired source task, with no skill from the target task. Baseline uses no skills.

Setting	Final SR $\uparrow$	1st SR $\uparrow$	SR@30m $\uparrow$	Avg. Ep. $\downarrow$	Avg. Time (min) $\downarrow$
Baseline	48.0%	12.0%	4.0%	3.50	180.8
$\mathit{RoboSkill}^{-}$ from Astra	80.0%	58.0%	20.0%	1.33	101.2
$\mathit{RoboSkill}^{-}$ from Sol	84.0%	42.0%	32.0%	3.56	86.5
$\mathit{RoboSkill}^{-}$ from Fable	92.0%	66.0%	60.0%	2.50	54.9
$\mathit{RoboSkill}^{-}$ from Opus	94.0%	58.0%	48.0%	1.84	52.4

Table 5 : Cross-agent skill transfer to Kimi K3 on LIBERO-10. We evaluate 5 seeds per task, totaling 50 cells per setting.

Setting	1st SR $\uparrow$	Avg. Time (min) $\downarrow$
Baseline	13.3%	34.8
Skills from Astra	70.0%	30.5

Table 6 : Cross-agent transfer on real-world easy tasks. Skills transfer from GPT-6 Astra to GPT-5.6 Sol.

GPT-5.6 Sol
Model / setting	Final SR $\uparrow$	1st SR $\uparrow$	SR@30m $\uparrow$	Avg. Ep. $\downarrow$	Avg. Time (min) $\downarrow$
Baseline	86.0%	30.0%	38.0%	9.0	80.3
$\mathit{RoboSkill}^{-}$	96.0%	58.0%	66.0%	3.7	47.1
$\mathit{RoboSkill}^{-}$ (1st update)	94.0%	58.0%	68.0%	3.7	49.3
$\mathit{RoboSkill}^{-}$ (2nd update)	98.0%	70.0%	82.0%	3.1	30.2
Opus 5

Table 7 : Iterative skill refinement on LIBERO-10. Baseline uses no saved skills.

\mathit{RoboSkill}^{-}

uses an initial skill constructed from successful exploration on the target task. Each update revises that skill through further successful exploration by the same model.

Setting	1st SR $\uparrow$	Avg. Time (min) $\downarrow$
Baseline	80.0%	27.0
$\mathit{RoboSkill}^{-}$	88.3%	23.1
$\mathit{RoboSkill}^{-}$ (1st update)	93.3%	26.2

Table 8 : Skill evolution on real-world hard tasks using GPT-6 Astra.

	GPT-6 Astra		GPT-5.6 Sol		Fable 5.1		Opus 5
Setting	1st SR $\uparrow$	Avg. Time (min) $\downarrow$	1st SR $\uparrow$	Avg. Time (min) $\downarrow$	1st SR $\uparrow$	Avg. Time (min) $\downarrow$	1st SR $\uparrow$	Avg. Time (min) $\downarrow$
$\mathit{RoboSkill}^{-}$ (w/o Code)	90.0%	21.1	15.0%	67.1	90.0%	16.1	70.0%	32.1
$\mathit{RoboSkill}^{-}$	92.5%	16.5	67.5%	39.9	90.0%	13.0	67.5%	30.4

Table 9 : Effect of executable code on LIBERO-10. The w/o Code variant removes code from the skill provided to

\mathit{RoboSkill}^{-}

while retaining its textual guidance.

	LIBERO-10					Real Robot
Setting	Final SR $\uparrow$	1st SR $\uparrow$	SR@30m $\uparrow$	Avg. Ep. $\downarrow$	Avg. Time (min) $\downarrow$	1st SR $\uparrow$	Avg. Time (min) $\downarrow$
With tactile	100%	78%	94%	1.42	15.93	90%	20.00
w.o. tactile	100%	82%	74%	1.20	25.35	80%	33.63

Table 10 : Importance of tactile input during exploration using GPT-6 Astra. In simulation, we evaluate 5 seeds per task, totaling 50 cells per setting.

Figure 3 : Real-robot hardware setup. Only the right arm is used in the experiments, together with its wrist-mounted camera and tactile-equipped gripper. The left arm is shown but is not used. The upper annotation indicates the support for the overhead third-person camera, which is outside the photographed field of view. The lower panel illustrates the tactile sensor, its sensing surface, and an example tactile image.

4 seeds per task: 40 cells
Model	Final SR $\uparrow$	1st SR $\uparrow$	SR@30m $\uparrow$	Avg. Ep. $\downarrow$	Avg. Time (min) $\downarrow$
GPT-6 Astra	100.0%	72.5%	87.5%	1.50	18.48
GPT-5.6 Sol	82.5%	27.5%	32.5%	9.20	94.50
Fable 5.1	97.5%	85.0%	77.5%	2.53	34.80
Opus 5	100.0%	60.0%	17.5%	1.93	80.90
5 seeds per task: 50 cells

Table 11 : Baseline results under the two seed protocols. The 4-seed protocol excludes the skill-construction seed and use seed 1-4 for each task; the 5-seed protocol includes all seeds 0–4.

Figure 4 : Example initial states of the ten LIBERO-10 tasks. Task IDs are used consistently in the cross-task pairings and skill selection tables. Each image illustrates one initial state.

ID	Task	Objective
Easy tasks
(a)	Press button	Approach and press the button to activate it.
(b)	Lift paper bag	Grasp a handle and lift the bag from the table.
(c)	Open drawer	Grasp the drawer and pull it outward.
(d)	Remove test tube	Grasp a test tube and extract it from the rack.
(e)	Sort socks	Arrange white socks on the left and black socks on the right.

Figure 5 : Execution stages of the 12 reported real-robot tasks. Panels (a)–(f) show easy tasks and panels (g)–(l) show hard tasks. Task objectives are listed in table 12 .

ID	Task	Objective
Easy tasks
(a)	Press button	Approach and press the button to activate it.
(b)	Lift paper bag	Grasp a handle and lift the bag from the table.
(c)	Open drawer	Grasp the drawer and pull it outward.
(d)	Remove test tube	Grasp a test tube and extract it from the rack.
(e)	Sort socks	Arrange white socks on the left and black socks on the right.

Table 12 : Real-robot task objectives. Letters correspond to the panels in figure 5 .

Easy tasks
Task	Baseline		$\mathit{RoboSkill}$
	Final SR $\uparrow$	Avg. Time (min) $\downarrow$	Final SR $\uparrow$	Avg. Time (min) $\downarrow$
Press button	10/10	17.68	10/10	9.10
Lift paper bag	10/10	19.58	10/10	15.70
Open drawer	8/10	22.53	10/10	15.40
Remove test tube	9/10	20.00	10/10	16.00

Table 13 : Task-wise Astra performance before and after exploration. Baseline timing uses the separate successful-sample records. Setting: 12-task subset (six easy and six hard), ten trials per task for success rates, 1 h per trial.

Task	Baseline		$\mathit{RoboSkill}$ from Astra
	Final SR $\uparrow$	Avg. Time (min) $\downarrow$	Final SR $\uparrow$	Avg. Time (min) $\downarrow$
Press button	0/5	–	5/5	25.40
Lift paper bag	2/5	27.00	5/5	19.60
Open drawer	1/5	27.00	3/5	34.33
Remove test tube	1/5	58.00	3/5	35.67
Sort white socks left and black socks right	0/5	–	2/5	29.00

Table 14 : Task-wise GPT-5.6 Sol performance before and after receiving skills constructed by Astra. Avg. Time uses successful trials only. Setting: six easy tasks, five trials per task, tactile input, and a 1 h limit per trial.

Target task	Rule	GPT-5.6 Sol	Opus 5	Fable 5.1	GPT-6 Astra
T0: Two cans to basket	[T0,T7,T1]	[T0,T7,T1]	[T0,T7,T1]	[T0,T7,T1]	[T0,T7]
T1: Cheese and butter to basket	[T1,T7,T0]	[T1,T7]	[T1,T7,T0]	[T1,T7,T0]	[T1,T7]
T2: Turn on stove; place moka pot	[T2,T8]	[T2,T8]	[T2,T8]	[T2,T8]	[T2,T8]
T3: Black bowl to drawer; close	[T3,T9]	[T3,T9,T5]	[T3,T9]	[T3,T9]	[T3]
T4: Two cups to left/right plates	[T4,T6,T9]	[T4,T6]	[T4,T6]	[T4,T6]	[T4,T6]
T5: Book to rear caddy slot	[T5]	[T5,T3]	[T5,T7]	[T5,T3]	[T5]

Table 15 : Agent-selected versus rule-based Top-K skill sets. Each cell lists the source-task skills selected for the target task. Bold agent-selected cells differ from the fixed rule in membership or order.

Model / setting	Final SR $\uparrow$	1st SR $\uparrow$	SR@30m $\uparrow$	Avg. Ep. $\downarrow$	Avg. Time (min) $\downarrow$
GPT-6 Astra
Rule-based Top-K	100%	87.5%	85%	1.175	19.2
Agent-selected Top-K	100%	97.5%	95%	1.025	17.1
$\Delta$ agent vs. rule	+0 pp	+10 pp	+10 pp	-12.8%	-11.2%
GPT-5.6 Sol
Rule-based Top-K	97.5%	57.5%	70%	3.425	35.4

Table 16 : Rule-based versus agent-selected Top-K performance.

\Delta

is the change from rule-based to agent-selected Top-K (percentage points for rates and relative percent otherwise). Bold marks the better non-tied value within each model and metric.

Explore similar work

ASPIRE: Agentic /Skills Discovery for Robotics
Jun 30, 2026Runyu Lu, Yubo Wu, Ethan Kou +11Human-To-Robot Transfer Robotics
EmbodiedSkills: A Unified Framework for Orchestrating, Training, and Deploying VLA Agents
Sep 1, 2026Wei Wang, Wenqiao Zhang, Yutong Lin +14Flow-Matching Vision-Language-Action Embodied Agents
Playful Agentic Robot Learning
Jun 17, 2026Junyi Zhang, Jiaxin Ge, Hanjun Yoo +17Scalable Robot Learning Agentic Reinforcement-Learning Framework

Jun 30, 2026, cs.ROJ/K move · Enter open · S save

ASPIRE: Agentic /Skills Discovery for Robotics

Runyu Lu, Yubo Wu, Ethan Kou, Letian Fu+10

NVIDIA · UMich · UIUC+2

Human-To-Robot Transfer Robotics Social Robot Navigation Continual Learning Methods

Open paper AskPDF arXiv

Traditional robot programming is challenging: it requires orchestrating multimodal perception, managing physical contact dynamics, and handling diverse configurations and execution failures. We introduce ASPIRE (Agentic Skill Programming through Iterative Robot Exploration), a continual learning system that autonomously writes and refines robot control programs in a code-as-policy paradigm while compounding experience into a reusable skill library. ASPIRE discovers skills that persist across tasks, simulation and real-world settings, and embodiments. It operates in an open-ended loop with three components: (1) a closed-loop robot execution engine that exposes fine-grained multimodal traces, enabling autonomous failure diagnosis, repair synthesis, and validation; (2) a continually expanding skill library that distills validated fixes into reusable, transferable knowledge; and (3) evolutionary search that generates diverse task sequences and control programs to explore beyond single-trajectory refinement. ASPIRE surpasses prior methods by up to 77% on LIBERO-Pro manipulation under perturbation, 72% on Robosuite bimanual handover, and 32% on BEHAVIOR-1K long-horizon household tasks. Its accumulated library also enables zero-shot generalization to unseen long-horizon tasks: on LIBERO-Pro Long, ASPIRE achieves 31% success versus 4% for prior methods despite their use of test-time reasoning and retries. Finally, simulation-discovered skills provide initial evidence of sim-to-real transfer, substantially reducing real-robot programming effort across different embodiments and robot APIs.

Abstract

Explore similar work

ASPIRE: Agentic /Skills Discovery for Robotics

EmbodiedSkills: A Unified Framework for Orchestrating, Training, and Deploying VLA Agents

Playful Agentic Robot Learning