--- title: REFACTOR-VLA: Unsupervised Library Learning of Typed... description: Apple researchers introduce REFACTOR-VLA, a robotics system that discovers reusable motor skills using a wake/sleep architecture. During sleep, a... canonical: https://daily.dev/posts/refactor-vla-unsupervised-library-learning-of-typed-motor-programs-jfp3dhys6 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: REFACTOR-VLA: Unsupervised Library Learning of Typed Motor Programs | daily.dev og:description: Apple researchers introduce REFACTOR-VLA, a robotics system that discovers reusable motor skills using a wake/sleep architecture. During sleep, a... og:url: https://daily.dev/posts/refactor-vla-unsupervised-library-learning-of-typed-motor-programs-jfp3dhys6 og:image: https://api.daily.dev/og/posts/Jfp3dhYS6.png og:image:alt: REFACTOR-VLA: Unsupervised Library Learning of Typed Motor Programs og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # REFACTOR-VLA: Unsupervised Library Learning of Typed Motor Programs **[Apple Machine Learning Research](https://daily.dev/sources/mlapple)** · 2 min read · 0 upvotes · 0 comments ## Summary Apple researchers introduce REFACTOR-VLA, a robotics system that discovers reusable motor skills using a wake/sleep architecture. During sleep, a Behavioral-Equivalence Kernel clusters motor program segments based on rollouts in a learned latent world model. During wake, typed lambda terms drawn from a Hindley-Milner-inspired vocabulary are generated and fed to a library-conditioned rectified-flow action decoder. Abstractions must pass a Minimum Description Length criterion and a return-preservation gate to become skills. Evaluated on the LIBERO benchmark suite, results show that simply scaling the world model from 188M to 430M parameters hurt performance across all four suites, while adding an InfoNCE contrastive loss during world-model warmup substantially improved skill clustering quality, measured via Normalized Mutual Information across the Object, Spatial, Goal, and LIBERO-10 suites. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### Does making the world model bigger improve performance in REFACTOR-VLA's robot skill learning system? No, increasing the latent world model size from 188 million to 430 million parameters worsened performance on all 4 of the LIBERO benchmark suites tested in the REFACTOR-VLA system, disproving the assumption that scaling up a world model always helps skill discovery quality. _Robotics researchers weighing model scaling tradeoffs can track findings like this on daily.dev._ ### What is a Behavioral-Equivalence Kernel in robot skill discovery? A Behavioral-Equivalence Kernel (BEK) is a method used in REFACTOR-VLA to determine whether two robot action sequences are behaviorally equivalent, based on rolling out those actions in a learned latent world model rather than clustering contrastive embeddings or relying on an uncalibrated LLM judgment, as prior approaches like AtomicVLA, BLADE, and LRLL do. _Engineers comparing skill-clustering approaches for robot policies can follow this research on daily.dev._ ### Does adding an InfoNCE contrastive loss improve skill clustering in robot world models? Yes, adding an auxiliary supervised contrastive InfoNCE loss during world-model warmup significantly improved skill clustering quality in the sleep phase of REFACTOR-VLA. Measured via Normalized Mutual Information across LIBERO suites, scores reached 0.462 (Object), 0.867 (Spatial), 0.915 (Goal), and 0.754 (LIBERO-10) under 3-seed evaluation. _Teams tuning contrastive objectives for representation learning can follow developments like this on daily.dev._ ## Similar posts on daily.dev - [LatentVLA: Latent Reasoning Models for Autonomous Driving](https://daily.dev/posts/latentvla-latent-reasoning-models-for-autonomous-driving-dbnpsb6n6) · Towards Data Science · 1 upvotes · 0 comments - [A Comprehensive Overview of Vision-Language-Action Models](https://daily.dev/posts/a-comprehensive-overview-of-vision-language-action-models-3qx6p4llr) · DigitalOcean Community · 1 upvotes · 0 comments - [How Visual-Language-Action \(VLA\) Models Work](https://daily.dev/posts/how-visual-language-action-vla-models-work-jln0xmknn) · Towards Data Science · 2 upvotes · 0 comments --- Tags: [#apple](https://daily.dev/tags/apple), [#robotics](https://daily.dev/tags/robotics), [#world-models](https://daily.dev/tags/world-models) [View this post on daily.dev](https://daily.dev/posts/refactor-vla-unsupervised-library-learning-of-typed-motor-programs-jfp3dhys6) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"REFACTOR-VLA: Unsupervised Library Learning of Typed Motor Programs","url":"https://daily.dev/posts/refactor-vla-unsupervised-library-learning-of-typed-motor-programs-jfp3dhys6","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/refactor-vla-unsupervised-library-learning-of-typed-motor-programs-jfp3dhys6"},"datePublished":"2026-09-02T15:43:10.374Z","dateModified":"2026-09-02T16:10:13.314Z","description":"Apple researchers introduce REFACTOR-VLA, a robotics system that discovers reusable motor skills using a wake/sleep architecture. During sleep, a...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/8dda20dea5f9a8a0e25a1b7bb36a1a20?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/8dda20dea5f9a8a0e25a1b7bb36a1a20?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Apple Machine Learning Research","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Apple Machine Learning Research","logo":"https://media.daily.dev/image/upload/s--mD-1uSe0--/f_auto,q_auto/v1785654191/logos/mlapple","url":"https://daily.dev/sources/mlapple"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/refactor-vla-unsupervised-library-learning-of-typed-motor-programs-jfp3dhys6","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"apple,robotics,world-models","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Apple Machine Learning Research","item":"https://daily.dev/sources/mlapple"},{"@type":"ListItem","position":3,"name":"REFACTOR-VLA: Unsupervised Library Learning of Typed Motor Programs"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/refactor-vla-unsupervised-library-learning-of-typed-motor-programs-jfp3dhys6#faq","mainEntity":[{"@type":"Question","name":"Does making the world model bigger improve performance in REFACTOR-VLA's robot skill learning system?","acceptedAnswer":{"@type":"Answer","text":"No, increasing the latent world model size from 188 million to 430 million parameters worsened performance on all 4 of the LIBERO benchmark suites tested in the REFACTOR-VLA system, disproving the assumption that scaling up a world model always helps skill discovery quality. Robotics researchers weighing model scaling tradeoffs can track findings like this on daily.dev."}},{"@type":"Question","name":"What is a Behavioral-Equivalence Kernel in robot skill discovery?","acceptedAnswer":{"@type":"Answer","text":"A Behavioral-Equivalence Kernel (BEK) is a method used in REFACTOR-VLA to determine whether two robot action sequences are behaviorally equivalent, based on rolling out those actions in a learned latent world model rather than clustering contrastive embeddings or relying on an uncalibrated LLM judgment, as prior approaches like AtomicVLA, BLADE, and LRLL do. Engineers comparing skill-clustering approaches for robot policies can follow this research on daily.dev."}},{"@type":"Question","name":"Does adding an InfoNCE contrastive loss improve skill clustering in robot world models?","acceptedAnswer":{"@type":"Answer","text":"Yes, adding an auxiliary supervised contrastive InfoNCE loss during world-model warmup significantly improved skill clustering quality in the sleep phase of REFACTOR-VLA. Measured via Normalized Mutual Information across LIBERO suites, scores reached 0.462 (Object), 0.867 (Spatial), 0.915 (Goal), and 0.754 (LIBERO-10) under 3-seed evaluation. Teams tuning contrastive objectives for representation learning can follow developments like this on daily.dev."}}]} ```