The Duality of Generative AI and Reinforcement Learning in Robotics: A Review

• tag:aim-pro.eu,2026:oer/5835121c01b5
• https://arxiv.org/abs/2410.16411
• arxiv:2410.16411
• doi:10.1016/j.inffus.2025.104003

This engine's id, the source URI, and any persistent identifier the source published.

• The Duality of Generative AI and Reinforcement Learning in Robotics: A Review

The title of this resource, not of the record or repository it came from.

• dcmitype:Text
• preprint

A DCMI Type term for the form held, plus the source's own genre label as declared.

• Angelo Moroncelli
• Vishal Soni
• Marco Forgione
• Dario Piga
• Blerina Spahiu
• Loris Roveda

Only agents the source named as creators. A repository owner is not made an author by default.

• Recently, generative AI and reinforcement learning (RL) have been redefining what is possible for AI agents that take information flows as input and produce intelligent behavior. As a result, we are seeing similar advancements in embodied AI and robotics for control policy generation. Our review paper examines the integration of generative AI models with RL to advance robotics. Our primary focus is on the duality between generative AI and RL for robotics downstream tasks. Specifically, we investigate: (1) The role of prominent generative AI tools as modular priors for multi-modal input fusion in RL tasks. (2) How RL can train, fine-tune and distill generative models for policy generation, such as VLA models, similarly to RL applications in large language models. We then propose a new taxonomy based on a considerable amount of selected papers. Lastly, we identify open challenges accounting for model scalability, adaptation and grounding, giving recommendations and insights on future research directions. We reflect on which generative AI models best fit the RL tasks and why. On the other side, we reflect on important issues inherent to RL-enhanced generative policies, such as safety concerns and failure modes, and what are the limitations of current methods. A curated collection of relevant research papers is maintained on our GitHub repository, serving as a resource for ongoing research and development in this field: https://github.com/clmoro/Robotics-RL-FMs-Integration.

The abstract, and the depositor's additional notes after it when the source has a field for them.

• cs.RO (arxiv)
• cs.LG (arxiv)

Declared keywords and taxonomy terms, each carrying its scheme.

• en

The language of the content as the language gate read it.

• http://creativecommons.org/licenses/by/4.0/

The licence document, as a URI where the gate resolved one.

• CC-BY-4.0 — assessed as OPEN by the harvester's licence gate. Conditions: attribution required. open, permissive

The verdict and the conditions, in words.

• open

The access condition the source declared. Not the licence.

• text/markdown

One value per form held — the original and the extracted text.

• 127545 bytes (extracted text)

Sizes in bytes, each saying which form it measures.

• 2024-10-21

When the resource was published. Not when it was collected.

• 2025-07-18

When the source's own record was last changed.

• Information Fusion Volume 129, May 2026, 104003

Where the work was published, in the source's own words: a journal with its volume and pages, a conference, an imprint.

• arXiv:2303.08774
• arXiv:2308.12270
• arXiv:2204.01691
• arXiv:2211.15657
• arXiv:2312.09187
• arXiv:2301.08028
• arXiv:2402.08546
• arXiv:2309.13041
• arXiv:2108.07258
• arXiv:2212.06817
• arXiv:2307.15818
• arXiv:2404.00282
• and 86 more, every one of them in the export

Works this one cites, when the source declares them as relations. What its text links to and its reference list cites is inferred, and stands under it apart.

• Retrieved from arXiv on 2026-10-09 in response to the search string “(all:"artificial intelligence" OR all:"machine learning" OR all:"generative AI" OR all:"deep learning" OR all:"reinforcement learning" OR all:"large language model") AND (all:"AI concepts" OR all:"types of AI" OR all:"AI fundamentals" OR all:"recognizing AI" OR all:"recognising AI" OR all:"general versus narrow AI" OR all:"narrow AI" OR all:"general AI" OR all:"machine intelligence" OR all:"AI strengths and weaknesses" OR all:"traditional software" OR all:"rule-based systems" OR all:"introduction to AI" OR all:"introduction to artificial intelligence" OR all:"artificial intelligence introduction" OR all:"AI primer" OR all:"foundations of artificial intelligence" OR all:"overview of AI" OR all:"understanding AI" OR all:"history of AI" OR all:"AI essentials" OR all:"AI terminology" OR all:"metaphors for AI" OR all:"AI fundamental concepts" OR all:"AI key concepts" OR all:"philosophy of AI" OR all:"critical AI literacy")”. arXiv served the resource and is not asserted to be its publisher or author.
• Text extracted from latexml-html to Markdown by arxiv-html; the original is retained unchanged beside it.

Where it was collected from, what was converted, and what container it came out of — the custody statements that would otherwise be mistaken for authorship.

• URI: tag:aim-pro.eu,2026:oer/5835121c01b5
• URI: https://arxiv.org/abs/2410.16411
• ARXIV: 2410.16411
• DOI: 10.1016/j.inffus.2025.104003

This engine's identifier for the resource, the source URI, and any persistent identifier the source published.

• The Duality of Generative AI and Reinforcement Learning in Robotics: A Review

• en

The language of the content, as the language gate read it off the converted text. The language the source declared is kept separately.

• Recently, generative AI and reinforcement learning (RL) have been redefining what is possible for AI agents that take information flows as input and produce intelligent behavior. As a result, we are seeing similar advancements in embodied AI and robotics for control policy generation. Our review paper examines the integration of generative AI models with RL to advance robotics. Our primary focus is on the duality between generative AI and RL for robotics downstream tasks. Specifically, we investigate: (1) The role of prominent generative AI tools as modular priors for multi-modal input fusion in RL tasks. (2) How RL can train, fine-tune and distill generative models for policy generation, such as VLA models, similarly to RL applications in large language models. We then propose a new taxonomy based on a considerable amount of selected papers. Lastly, we identify open challenges accounting for model scalability, adaptation and grounding, giving recommendations and insights on future research directions. We reflect on which generative AI models best fit the RL tasks and why. On the other side, we reflect on important issues inherent to RL-enhanced generative policies, such as safety concerns and failure modes, and what are the limitations of current methods. A curated collection of relevant research papers is maintained on our GitHub repository, serving as a resource for ongoing research and development in this field: https://github.com/clmoro/Robotics-RL-FMs-Integration.

The abstract, and the depositor's additional notes after it as a second LangString when the source has a field for them.

• cs.RO
• cs.LG

• v2

• revised

Draft, final, revised or unavailable, and only when the source says so. A first version is not thereby final.

• Angelo Moroncelli — author
• Vishal Soni — author
• Marco Forgione — author
• Dario Piga — author
• Blerina Spahiu — author
• Loris Roveda — author

Role, entity and date per declared contribution. A role outside LOM's vocabulary is reported in the entry's description instead.

• URI: tag:aim-pro.eu,2026:oer/5835121c01b5/record

The identifier of this metadata record — the resource's own, with /record after it, because the record is a description of the resource and not the resource.

• AIM-PRO WP3 OER harvester (arXiv) — creator

Who generated this record and when — a statement about the record, not about the resource.

• LOMv1.0
• aimpro-oer-profile/1

LOMv1.0, and the profile this was built against.

• en

• text/markdown

One value per form held: the original as the source published it, and the Markdown this engine extracted.

• 127545

Bytes. The original's, because the resource is the file and not our conversion of it.

• https://arxiv.org/abs/2410.16411

Where the source serves it.

• narrative text

Only from a declaration that maps onto LOM's vocabulary. A file format is not a declaration.

• no — free to use

No: the licence gate refuses anything it cannot establish open reuse terms for, before the bytes are requested.

• yes — attribution required (CC-BY-4.0)

Yes unless the licence reserves nothing — attribution is a restriction. The conditions after the dash are the licence gate's reading; the export carries LOM's bare term.

• CC-BY-4.0 · http://creativecommons.org/licenses/by/4.0/. Conditions: adaptation permitted, attribution required.

The SPDX id, the licence URI and the conditions. LOM has no element for any of the three, so this is where they survive.

• ispartof

• Information Fusion Volume 129, May 2026, 104003

• references: arXiv:2303.08774
• references: arXiv:2308.12270
• references: arXiv:2204.01691
• references: arXiv:2211.15657
• references: arXiv:2312.09187
• references: arXiv:2301.08028
• references: arXiv:2402.08546
• references: arXiv:2309.13041
• references: arXiv:2108.07258
• references: arXiv:2212.06817
• references: arXiv:2307.15818
• references: arXiv:2404.00282
• and 86 more, every one of them in the export

The container a file was found inside, and the Markdown extracted from the original. What a lab requires, and the lab a component belongs to, are inferred and stand apart.