Autonomous Skill Acquisition (ASA)

• Autonomous Skill Acquisition (ASA) is a collection of methods where agents progressively build modular, reusable procedural skills for various tasks.
• It integrates frameworks like goal-conditioned reinforcement learning, repository mining, and LLM-driven externalization to map goals to effective policies.
• ASA approaches emphasize safety, verification, and lifecycle management to evolve skill repertoires that adapt to dynamic environments.

Searching arXiv for the primary paper and closely related ASA work to ground the article in current literature. Autonomous Skill Acquisition (ASA) denotes a family of methods in which an agent, robot, or learner progressively acquires reusable procedural competence rather than solving each task from scratch. In the current literature, the term covers several related regimes: open-ended goal generation in reward-free or multi-goal reinforcement learning, discovery of parameterized skills and their applicability regions in robotics, externalization of reusable skills as explicit artifacts for LLM agents, automated mining of procedural knowledge from repositories and interaction traces, and assisted human or expert-in-the-loop skill formation (Srivastava et al., 6 Feb 2025, Cartoni et al., 2018, Bi et al., 12 Mar 2026, Wang et al., 9 Apr 2025, Richards et al., 30 Jun 2025). Across these settings, the central shift is from monolithic competence to modular, accumulable, and often testable skills.

1. Formalizations and core abstractions

ASA has no single formalization, but the literature converges on a small set of recurring abstractions: a skill has internal procedure, a boundary for invocation, and some notion of applicability or goal-conditioning. In repository-mined agent systems, a skill is defined as a four-tuple

$S = (\mathcal{C}, \pi, \mathcal{T}, \mathcal{R}),$

where $\mathcal{C}$ is the applicability condition, $\pi$ the policy or procedural sequence, $\mathcal{T}$ the termination criteria, and $\mathcal{R}$ the interface (Bi et al., 12 Mar 2026). In parameterized-skill robotics, the central object is instead a mapping from goals to policy parameters,

$\Theta: O \rightarrow \Psi,$

with $O$ an achievable outcome space and $\Psi$ a policy-parameter space (Cartoni et al., 2018). In option-style reinforcement learning, ASAP defines a skill as a temporally extended action $\zeta=\langle \sigma_\theta, p\rangle$, with $\sigma_\theta$ an intra-skill policy and $\mathcal{C}$0 a termination distribution (Mankowitz et al., 2016). In LLM personalization systems, skills are represented as explicit, versioned artifacts such as

$\mathcal{C}$1

where the fields encode name, description, prompt, triggers, tags, examples, and version (Yang et al., 1 Mar 2026).

Source	Skill representation	Emphasis
(Bi et al., 12 Mar 2026)	$\mathcal{C}$2	Applicability, procedure, termination, interface
(Cartoni et al., 2018)	$\mathcal{C}$3	Goal-to-policy mapping
(Mankowitz et al., 2016)	$\mathcal{C}$4	Option-like skill with termination
(Yang et al., 1 Mar 2026)	$\mathcal{C}$5	Explicit, versioned artifact
(Wu et al., 3 Feb 2026)	$\mathcal{C}$6	Instruction Injection, Scoped Tools, Hard Constraints

A major conceptual distinction in the recent agent literature is between a skill and a mere tool wrapper. The four-tuple definition makes this explicit: a skill includes contextual activation rules, embedded reasoning or procedure, success conditions, and a standardized callable boundary, whereas a tool wrapper exposes only an operation (Bi et al., 12 Mar 2026). Related work on MUSE-Autoskill and AutoSkill pushes the same distinction further by treating skills as long-lived assets with memory, validation, and versioning rather than transient prompt fragments (Lin et al., 26 May 2026, Yang et al., 1 Mar 2026). This suggests that ASA increasingly concerns not only learning procedures, but also stabilizing them as reusable infrastructure.

2. Open-ended goal discovery and skill formation in reinforcement learning and robotics

A foundational ASA line treats skill acquisition as the autonomous construction of goal-conditioned behavior. Autotelic reinforcement learning frames the problem in a reward-free MDP in which agents must autonomously represent, generate, and master their own goals. The relevant policy template is goal-conditioned,

$\mathcal{C}$7

and intrinsic motivation is divided into knowledge-based and competence-based forms: the former drives novelty-seeking and exploration, while the latter favors goals at the frontier of current ability and thereby induces a developmental curriculum (Srivastava et al., 6 Feb 2025). IMGEPs operationalize this through a loop of goal generation, goal selection, attempted achievement, evaluation of progress, and update of future goal sampling.

When the goal space is unknown, ASA becomes a manifold-discovery problem. AGME assumes that achievable outcomes occupy a small subset $\mathcal{C}$8 of a much larger sensory space, and it avoids the curse of dimensionality by expanding outward from known achievable outcomes rather than sampling the full sensor space. It constructs a $\mathcal{C}$9-neighbor graph over achieved outcomes, scores frontier points by average neighbor distance,

$\pi$0

selects

$\pi$1

and perturbs the associated policy parameters with Gaussian noise to discover nearby outcomes (Cartoni et al., 2018). The resulting parameterized skill is implicit in the repertoire of achieved outcomes and associated policies.

Learned goal spaces extend this idea to real visuomotor settings. A 6-joint robotic arm manipulating a ball in an arena can learn goals in a VAE latent space rather than a hand-crafted feature space, and then use an inverse-model meta-policy with 1-nearest-neighbor retrieval to map latent goals to DMP parameters (Laversanne-Finot et al., 2019). ASAP addresses a complementary problem: not only learning skills, but learning where to apply them. It represents skill partitions with hyperplanes, defines a mixture policy

$\pi$2

and proves convergence to a local optimum under standard conditions (Mankowitz et al., 2016). The emphasis here is that reusable skill repertoires require both internal control and an inter-skill policy.

Real-world unsupervised ASA has recently been pushed further by URSA. URSA extends QDAC with a cost critic $\pi$3, a fixed-size repertoire $\pi$4, and optional VAE-based unsupervised feature learning. It defines a skill as the expected feature vector under a policy’s stationary distribution,

$\pi$5

and optimizes return subject to skill-matching and safety constraints. On a Unitree A1 quadruped, URSA covers about $\pi$6 more of a discretized joint-angle space than the baselines and gives best performance in $\pi$7 simulated and $\pi$8 real-world damage scenarios (Grillotti et al., 26 Aug 2025). A plausible implication is that ASA in robotics is moving from handcrafted behavior descriptors toward learned, safety-constrained repertoires that remain reusable under damage and task shift.

3. Externalized skills and lifecycle management in LLM agents

In LLM-agent work, ASA is increasingly defined as post-deployment acquisition of executable expertise without retraining model weights. The repository-mining framework of "Automating Skill Acquisition through Large-Scale Mining of Open-Source Agentic Repositories" standardizes skills in the SKILL.md format and organizes them by progressive disclosure: Level 1 YAML frontmatter loaded at startup, Level 2 instructions loaded when triggered, and Level 3 resources loaded on demand (Bi et al., 12 Mar 2026). This architecture is intended to preserve a large skill library without bloating the context window.

Programmatic skill induction makes the acquired skill executable rather than descriptive. ASI induces Python functions from successful web trajectories, rewrites the original episode to force use of the newly induced functions, and accepts the skill only if replay succeeds, uses the new skill, and each skill-calling action produces environment changes. On WebArena, ASI reaches a success rate of 40.4 versus 32.7 for the static baseline and 36.3 for the text-skill baseline, while reducing average steps to 5.0 from 5.6 and 5.9 (Wang et al., 9 Apr 2025). The paper attributes the gain primarily to programmatic verification during induction, and the contrast with text-only memory suggests that ASA benefits from executable representations with explicit acceptance tests.

Lifecycle-managed skill systems push this further by treating skills as evolving assets. MUSE-Autoskill organizes skills through creation, memory, management, evaluation, and refinement inside a ReAct-style runtime loop. Skills are packaged with SKILL.md, optional scripts/, resources/, and tests/, and only pass into the skill bank if unit tests succeed in a sandbox. Each skill also carries a sibling .memory.md that accumulates failure modes, input quirks, caveats, and performance lessons (Lin et al., 26 May 2026). On SkillsBench, MUSE reports successful skill generation on 35/51 tasks, overall self-created-skill accuracy of 60.35%, and 87.94% Phase 2 accuracy on the 35 tasks where skills were generated (Lin et al., 26 May 2026).

Other systems formalize skill evolution at the artifact level. AutoSkill extracts reusable behavioral knowledge from user queries, stores it as explicit versioned skills, retrieves it by a hybrid dense-plus-BM25 relevance score, and updates the skill bank through add, merge, or discard decisions (Yang et al., 1 Mar 2026). EvoDS synthesizes executable tools $\pi$9, validates them through execution, caches them in a synthesized skill repository, and promotes them into a sub-agent’s permanent action space only if a usage count exceeds a threshold $\mathcal{T}$0 (Yang et al., 2 Jun 2026). Together, these systems replace the notion of “experience as prompt residue” with “experience as inspectable skill object.”

4. Automated acquisition from repositories, corpora, and interaction traces

A central recent development is the treatment of external ecosystems as ASA substrates. Repository mining exemplifies this most explicitly. The open-source ASA pipeline in (Bi et al., 12 Mar 2026) proceeds through repository structural analysis and contextualization, semantic skill identification through dense retrieval, and translation into SKILL.md. A repository-to-context tool such as repo2AI is used to expose directory trees and contents; a bi-encoder retrieves candidate modules by cosine similarity and a cross-encoder reranks them; modules above threshold $\mathcal{T}$1 are then translated into reusable skill packages. TheoremExplainAgent and Code2Video serve as case studies: the first yields a portable “visual theorem walkthrough” pattern, and the second yields a “visual layout critic” pattern with Visual Anchor Prompting and a $\mathcal{T}$2 grid (Bi et al., 12 Mar 2026). The reported result is a 40% gain in knowledge transfer efficiency for agent-generated educational videos relative to baseline code generation models, and the TEA o3-mini implementation achieves an overall score of 0.77 on TheoremExplainBench (Bi et al., 12 Mar 2026).

Automated context discovery addresses a related but distinct problem: not skill extraction from code, but identification of context that actually improves task performance. ECS searches over context units derived from external corpora and optimizes them by downstream accuracy on a development set rather than semantic similarity alone. On BackendBench, ECS reaches 0.461 average correctness and is reported as a 27% relative improvement over AST + Dense; on $\mathcal{T}$3-Bench airline it reaches Pass$\mathcal{T}$4=0.767 and improves the benchmark by 7% (Sun et al., 18 Feb 2026). Because the evolved contexts transfer from Gemini-3-Flash to Claude-4.5-Sonnet and DeepSeek-V3.2, the paper argues that task-useful context can itself function as an acquired skill substrate (Sun et al., 18 Feb 2026).

Scientific document automation provides a stronger example of corpus-grounded procedural knowledge acquisition. A modular skill-based agent centered on Qwen3-235B-A22B constructs a creep database through Literature Collection, Automated Screening, Multi-modal Information Extraction, Formula Validation, and Structured Storage (Wu et al., 3 Feb 2026). Its skills are defined as

$\mathcal{T}$5

where $\mathcal{T}$6 is Instruction Injection, $\mathcal{T}$7 Scoped Tools, and $\mathcal{T}$8 Hard Constraints. Applied to 243 publications, the system produces 353 high-fidelity equation-data pairs, exceeds 90% verified success for graphical data digitization, and uses cross-modal verification with a stated threshold $\mathcal{T}$9; in the AMAG-183 example, the agreement reaches $\mathcal{R}$0 (Wu et al., 3 Feb 2026). The authors explicitly characterize this as a precursor or instantiation of skill-based autonomous scientific automation rather than full ASA, because the skill modules are engineered rather than discovered (Wu et al., 3 Feb 2026).

A more indirect but important line studies ASA through the discovery of laws governing skill formation itself. The deep-to-symbolic pipeline in "Automated discovery of symbolic laws governing skill acquisition from naturally occurring data" fits a latent-state model to training logs, then uses symbolic regression to recover algebraic skill-learning laws (Liu et al., 2024). On Lumosity data, it reports new forms such as a logarithmic law for Reasoning,

$\mathcal{R}$1

and an inverse power-like form for Attention,

$\mathcal{R}$2

while obtaining the best BIC overall among the compared models (Liu et al., 2024). This broadens ASA from skill construction to scientific inference about the dynamics of skill acquisition.

5. Embodied, multimodal, and assisted skill acquisition

Robotic ASA has long included an embodied tradition in which a narrow seed behavior is generalized through play. In autonomous robotic playing, a user supplies a basic behavior $\mathcal{R}$3 valid only in a narrow domain of applicability, and the robot learns preparatory behaviors that transform new states into ones where the basic behavior succeeds. The environment model

$\mathcal{R}$4

is learned from play and used for active learning and creative generation of compound preparatory behaviors; in book grasping, active learning improves learning speed by around 9%, and creative composition enables previously unsolvable tasks such as tower disassembly (Hangl et al., 2017). The related skill-based programming paradigm combines this with visual programming, kinesthetic teaching, and skill-centric testing, so that users provide simple program skeletons while the robot generalizes them through autonomous playing and validates them through stored sensor traces and function-call profiles (Hangl et al., 2017).

Recent VLA work reframes ASA as continual acquisition of primitive-level control. InSight first segments demonstrations into labeled primitives through VLM plan decomposition and end-effector signals, then fine-tunes a steerable VLA that can execute primitive labels with a learned progress channel. For a new task, a VLM identifies any primitive gaps, parameterizes them as a single-axis motion in $\mathcal{R}$5, attempts autonomous collection, and retrains only on successful demonstrations (Wang et al., 23 Jun 2026). The system acquires block flipping, drawer closing, twisting, pouring, and sweeping without human demonstrations of those target skills; it reaches 75% success on block flipping after 246 acquired primitive rollouts and 479 total attempts, 100% success over 25 trials on drawer closing after retraining, 92% end-to-end success on twisting, 96% on pouring, 80% on a 14-primitive twist-then-pour composition, and 5/5 on sweeping while retaining 100% success on the original pick-and-place tasks (Wang et al., 23 Jun 2026).

Video-based dexterous ASA pushes autonomy further upstream. Video2Sim2Real starts from a single RGB-D human manipulation video, reconstructs a simulator-ready digital twin with Gemini, SAM3, and SAM3D, extracts robot and object motion priors with HaMeR, Mink, and CoTracker, refines behavior around object-centric contact, interaction, and detachment keyframes, then bridges sim-to-real via a PointNet-style IL policy for keyframe pose recalibration and a residual RL policy for local finger-level adaptations (Han et al., 7 Jun 2026). Across seven everyday manipulation tasks, the framework reaches 95.7% average real-world success, versus 15.7% for a FoundationPose-based adjustment baseline, 5.7% for Pure-IL, 45.7% for Pure-RL finger-only, and 3.0% for Pure-RL arm+finger (Han et al., 7 Jun 2026). The paper’s key claim is that the demonstrated skill should not be copied wholesale; instead, object-centric anchors should be used to preserve the demonstrated impact on the environment.

Not all embodied ASA is fully autonomous in the machine-learning sense. The tandem colonoscopy system is explicitly framed as expert-in-the-loop assisted skill acquisition rather than a robot learning a policy from demonstration. A custom preceptor control body telemanipulates the novice’s wheel controls in real time without handoff, using a binary switch and variable PID controller at 500 Hz, and switching to guidance mode when the expert rotates either wheel by more than 0.02 degrees between control cycles (Richards et al., 30 Jun 2025). In an exploratory study with 4 novice users and 1 expert, novices with expert intervention improved by 53.4% in time and 14.6% in path length over 10 trials, compared with 6.2% and 5.0% for novices without intervention (Richards et al., 30 Jun 2025). The broadened use of ASA here concerns preservation of in-hand motor synergies rather than autonomy of the machine.

Human skill formation under AI assistance adds a further cautionary dimension. In randomized Trio-programming experiments, AI assistance reduced post-task quiz scores by 17%, or about 2 grade points on a 27-point quiz, with Cohen’s $\mathcal{R}$6 and $\mathcal{R}$7, while delivering no significant average speedup (Shen et al., 28 Jan 2026). The study identifies six AI interaction patterns; three high-learning patterns preserve learning by maintaining conceptual engagement, whereas delegation-heavy patterns improve convenience at the cost of mastery (Shen et al., 28 Jan 2026). A plausible implication is that ASA must be evaluated differently depending on whether the acquiring system is the model, the robot, or the human using the model.

6. Evaluation, governance, and unresolved issues

ASA evaluation is intrinsically plural because the acquired object differs by paradigm. In autotelic RL, the literature emphasizes exploration diversity, unseen-goal generalization, and robustness under perturbation (Srivastava et al., 6 Feb 2025). In repository-mined educational agents, the evaluation dimensions are safety, completeness, executability, maintainability, and pedagogy, instantiated by vulnerability rate, feature coverage, success rate, schema drift, and TeachQuiz score (Bi et al., 12 Mar 2026). In lifecycle-managed agent skills, admission to the skill bank is mediated by unit tests and later updated by runtime feedback (Lin et al., 26 May 2026). In EvoDS, synthesized tools are retained only if they execute successfully and produce valid outputs, and subtask success is explicitly rewarded during GRPO-based training (Yang et al., 2 Jun 2026). These differences reflect a deeper fact: ASA metrics depend on whether the skill is a policy, a repertoire point, an executable tool, a document-processing module, or a human competence.

Security and governance have become central because many modern ASA systems acquire capabilities from external artifacts. The repository-mining study reports a vulnerability rate of 26.1% in a survey of community-distributed skills and proposes a four-stage governance pipeline: G1 static analysis for suspicious strings and destructive operations, G2 semantic classification for purpose alignment and hidden injections, G3 behavioral sandboxing with network and file isolation plus resource monitoring, and G4 permission validation against allowed-tools manifests (Bi et al., 12 Mar 2026). URSA encodes safety directly into the optimization problem through a discounted cost critic $\mathcal{R}$8 (Grillotti et al., 26 Aug 2025). InSight filters autonomous acquisitions by success and still requires human environment resets, while Video2Sim2Real explicitly measures safety and RMS jerk alongside success (Wang et al., 23 Jun 2026, Han et al., 7 Jun 2026). This suggests that scalable ASA increasingly resembles software supply-chain management as much as classical learning theory.

Several recurrent misconceptions are addressed directly in the literature. First, ASA is not identical to fine-tuning: many recent systems acquire new capabilities without updating base-model weights, instead using externalized skills, evolved context, or retrieved artifacts (Bi et al., 12 Mar 2026, Sun et al., 18 Feb 2026, Yang et al., 1 Mar 2026). Second, not every skill-based system constitutes full ASA in the strongest sense: some frameworks are better described as precursors or assisted variants because their skills are engineered or their control remains expert-mediated (Wu et al., 3 Feb 2026, Richards et al., 30 Jun 2025). Third, higher immediate task performance is not equivalent to genuine skill formation, especially when the learner is human rather than machine (Shen et al., 28 Jan 2026).

Taken together, the literature points toward a composable “agentic stack” in which skills provide procedural intelligence, external protocols provide tool connectivity, and evaluation combines executability, transfer, safety, and long-horizon maintainability (Bi et al., 12 Mar 2026). The open problems named across the cited work are consistent: schema drift and ontological organization as skill libraries scale, context distraction from overly large skill pools, improved promotion and pruning criteria, richer safety boundaries, better use of failed trajectories, stronger cross-agent transfer, and more faithful preservation of learning when AI acts as tutor rather than delegate (Lin et al., 26 May 2026, Yang et al., 2 Jun 2026, Grillotti et al., 26 Aug 2025, Shen et al., 28 Jan 2026). ASA is therefore less a single algorithm than a unifying research program on how procedural competence is discovered, represented, validated, and accumulated over time.