Generalist models outperform specialist policies in robot manipulation
Source published ·Modelwire updated
Original coverage: arXiv cs.LG ↗·How Modelwire adds context
The development
Generalist AI models are outperforming specialist robot control architectures in manipulation tasks, challenging the dominant paradigm of vision-language-action models. AMP1, a unified framework that adapts commands in real time across single-arm, bimanual, and mobile systems, achieves 85.6% zero-shot success on RoboLab-120 versus 49.2% for leading VLAs. This shift signals that end-to-end agentic control may scale more effectively than task-specific training pipelines, with implications for robotics deployment and the broader trend toward generalist foundation models replacing narrow specialist systems.
Modelwire’s AI-generated summary of coverage from arXiv cs.LG.
Modelwire analysis
Analyst takeOur AI-generated reading of the wider context and the next developments to watch.
The real story isn't that a generalist model beat specialist VLAs on one benchmark. It's that this result arrives just as the field is converging on foundation models for robotics (see the dexterous hands survey from Nov 4), suggesting the VLA paradigm may have been a local maximum rather than the path forward.
Two weeks ago, research showed generalist robot policies can self-improve across tasks during deployment without task-specific retraining. This AMP1 paper extends that finding into the architectural layer: end-to-end agentic control appears to scale better than the modular vision-language-action pipeline that dominated 2024-2025. The implication compounds with Baidu's world model work (Nov 4), which demonstrated that real sensorimotor data creates structural advantages. Together, these papers suggest companies betting on VLA-first robotics stacks may need to reconsider their foundation model strategy.
If Boston Dynamics, Sanctuary AI, or Figure AI announce architectural shifts away from vision-language-action frameworks toward end-to-end agentic control in the next six months, that confirms this is reshaping industry practice rather than remaining an academic curiosity. Conversely, if major robotics labs publish results showing VLAs match or exceed AMP1 on RoboLab-120 or comparable benchmarks by Q2 2027, the architectural question remains unsettled.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.LG
Foundation models reshape dexterous robotics across hardware and learning layers
A comprehensive survey maps the convergence of dexterous robotics and foundation models across five interconnected layers: hardware design, data pipelines, cross-embodiment transfer learning, model architectures, and evaluation frameworks. Covering January 2024 through November 2026, the work synthesizes how large-scale human motion datasets and pretrained representations are reshaping robot hand control, while tactile sensing and hand-object…
MentionsAMP1 · RoboLab-120 · CaP-Gym · vision-language-action models · world-action models
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.LG originally reported this story as “On Agentic Manipulation Policies”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.