The phrase "foundation model for robots" has moved from research aspiration to marketing copy, and 2026 finds the field crowded with systems that promise to translate a spoken instruction into a physical action. Much of what follows rests on first-party announcements and project documentation rather than independent replication, so the sections below separate what vendors and maintainers describe from what has been broadly verified. Where a claim has not been independently corroborated, it is presented as a claim, not a settled result.
Vision-Language-Action (VLA) Models: Foundation Models That Control Robot Actions
The umbrella term for this class of system is the vision–language–action (VLA) model. Reference material characterizes a VLA model as a type of model, developed in robot learning, that integrates vision, language, and action within a single framework — taking in images of the environment together with a natural-language instruction and producing robot actions as output (Wikipedia). Whether any given VLA constitutes a "foundation model that allows control of robot actions" in the strong sense — general, reusable, and reliable across robots and tasks — is not something the available sources establish independently; that framing is best read as a description of the design goal rather than a confirmed capability.
The conceptual appeal is straightforward. Rather than hand-engineering a perception pipeline, a planner, and a low-level controller separately, a VLA aims to learn the mapping from pixels-and-words to motor commands end to end. How well that mapping generalizes beyond the data a model was trained on remains an open empirical question, and readers should treat sweeping generalization claims with caution absent third-party evaluation.
What Today’s Best Robotics Foundation Models Can Do: Pick, Place, Sort, and Manipulate Objects from Natural Language
A common claim is that the best current robotics foundation models can follow natural-language instructions to pick, place, sort, and manipulate objects. This has not been independently cross-checked here and should be read as a claim rather than an established benchmark result. Demonstrations of language-conditioned manipulation do appear in vendor and research materials, but demo footage and controlled lab settings are weak evidence for robust, real-world performance.
That gap between demonstration and deployment is itself a recognized problem. NVIDIA has published guidance on how to evaluate general-purpose robot policies for real-world deployment, which frames evaluation — not just capability demos — as a central challenge for the field (NVIDIA Developer Blog). The existence of dedicated evaluation methodology is a useful signal: it implies the community does not yet treat "the robot picked up the object in the video" as sufficient proof of a reliable policy. Anyone assessing a "best model" claim should ask which policies were tested, on which embodiments, under what distribution shift, and with what success rates — details that headline capability statements typically omit.
Gemini Robotics ER 2: Video Understanding, Task Orchestration, and Multi-Robot Collaboration
Google DeepMind has announced Gemini Robotics ER 2, describing it in terms of video understanding, task orchestration, and multi-robot collaboration for robotics applications (DeepMind). These attributes are drawn from the announcement itself and have not been independently verified here; the framing throughout is the vendor’s own. Whether the described video-understanding, orchestration, and multi-robot features perform as characterized outside DeepMind’s own examples is not something the available evidence confirms.
The reason to flag this is not skepticism about the announcement’s good faith but the ordinary caution warranted by first-party product communications: a blog post describing what a system is designed to do is a statement of intent and capability by the party that built it, not an independent audit. Treat the three named capabilities as the vendor’s description of the product’s scope pending outside evaluation.
Post-Training NVIDIA Cosmos in One Day Using Agent Skills
An NVIDIA developer blog post presents a workflow for post-training NVIDIA Cosmos using agent skills, framed around doing so in roughly one day (NVIDIA Developer Blog). The specific claim — that Cosmos can be post-trained in one day using agent skills — has not been independently corroborated and should be read as NVIDIA’s own account of a workflow rather than a verified, generalizable result.
Several caveats attach to any "in one day" figure. Such timelines typically depend on hardware, dataset preparation that may not be counted in the headline number, a specific task definition, and a notion of "done" that may be narrower than a reader assumes. The available source establishes that NVIDIA describes this workflow; it does not establish that an arbitrary team would reproduce the same timeline or the same quality of result. The claim is therefore reported here as attributed, unverified, and conditional on details the summary does not fully specify.
Autonomous Coding Agents Pushing Vision Reasoning Models Above 90% Accuracy with Almost No Manual Effort
The proposition that autonomous coding agents can push vision reasoning models above 90% accuracy "with almost no manual effort" is, among the items covered here, the least substantiated. The sources available for this article do not confirm it, do not define the benchmark or dataset such a figure would refer to, and do not corroborate the "almost no manual effort" characterization. It is presented here only as an unverified claim.
Two specifics deserve scrutiny before any such number is taken at face value. First, "above 90% accuracy" is meaningless without a named benchmark, a task, and a baseline — the same percentage can be trivial on one dataset and state-of-the-art on another. Second, "almost no manual effort" tends to understate the human work embedded in dataset curation, evaluation design, prompt and harness construction, and the judgment of when a result is good enough to stop. Sources do not confirm any of these details, so this section makes no claim about whether the described outcome is achievable, and readers should not infer one.
openpi-vla-speedrun: A Data-Mixture Hill-Climb for Cross-Embodiment VLA Transfer on a Frozen π0.5 LoRA Trainer
The final item is an open-source project, openpi-vla-speedrun, hosted on GitHub (repository). The description of it as "a data-mixture hill-climb for cross-embodiment VLA transfer on a frozen π0.5 LoRA trainer" reflects how the project is characterized rather than an independently verified account of what it accomplishes; that specific framing has not been cross-checked here.
Each component of that phrase is a testable claim rather than a settled fact. A "data-mixture hill-climb" implies iteratively adjusting the proportions of training data to improve a metric; "cross-embodiment transfer" implies moving a policy across different robot bodies; a "frozen π0.5 LoRA trainer" implies fine-tuning a fixed base model with low-rank adaptation while the backbone stays unchanged. Whether the repository implements exactly this, and with what measured results, is something readers should confirm directly against the project’s own code, commits, and documentation rather than from a one-line summary. The link above is the appropriate starting point for that verification.
Across all six topics, the recurring pattern is that capability descriptions currently outrun independent confirmation. Vendor blogs, project READMEs, and reference articles establish what systems are claimed to do; they are a weaker basis for concluding what those systems reliably do in the wild. The evaluation-focused material on deploying robot policies is a reminder from within the field itself that the distance between a compelling demonstration and a dependable deployment is exactly where the hard, unresolved questions still sit.