What are VLAMs doing to Robotics?

Robotics is often described as a mature industry, with well-established actors, growing sales, and rather clear innovation directions and technological trajectories. For decades, robots have excelled at narrow, repetitive tasks in tightly controlled environments, but have yet failed in achieving complete flexibility and easy adaptation. Despite that, the “Mecca” of every actor in robotics remains the design of “generalist” robots with many affordances. Recent innovations allowing the convergence of robotics with AI seem on the right track to make the dream a bit closer. In our recent work, we argue that vision–language–action models (VLAMs) may mark a turning point in robotics. By integrating large language models (LLMs) with robotic perception and control, VLAMs introduce a new technological logic that has the potential to rejuvenate robotics as an industry, reshape where value is created, and alter who leads innovation. The increasing attention to embodied AI (the development of AI systems deeply rooted in hardware) is driven both by high expectations of future innovation and profitability, as well as from the awareness that the major trajectory of AI evolution — LLMs — taken alone, might not produce the promised and hyped economic returns.

At a high level, the most advanced VLAMs connect three modalities—vision, language, and action—within a single, end-to-end, learning architecture. This allows robots to interpret natural-language instructions, reason over visual inputs, and translate both into physical actions. In practical terms, VLAMs “bring problems of robotics into the semantic space”*: instead of programming robots through rigid task-specific code, developers can increasingly rely on language-based representations that generalise across tasks, environments, and embodiments.

A shift in how robots learn and adapt

The promise of VLAMs lies in flexibility. From a software angle, traditional robotic affordances have been constrained by data scarcity and fragmentation: every new task or environment typically requires bespoke datasets and extensive re-engineering. VLAMs offer one solution to partly bypass this bottleneck by leveraging LLMs trained on massive, general-purpose real-world datasets (as opposed to using AI to produce and train robotic systems on simulated data) that already encode rich information about objects, actions, and goals. When adapted to embodied settings, these models enable robots to recombine existing knowledge in novel ways—opening the door to more generalist, adaptable systems.

Conceptually, and following the conceptualisation introduced by Brian Arthur in his book “the Nature of Technology”, we interpret this shift as a case of redomaining. The core functions of robotics—perception, planning, and actuation—remain the same, but they are increasingly executed under a different technological principle. This makes VLAMs a radical innovation, even though they build on existing hardware and software components. And as it is well-known in the economics of technology, radical innovations are usually those that spur structural transformations, shifts the balance between winners and losers in a market as well as in the locus of value creation and capture, and sometimes even kickstart the emergence of brand-new markets. We witness such dynamics under our very own eyes with the fast-paced growth of the humanoid robot segment (despite the uncertainty still characterising its long-term prospects). Hence, forming an idea about the potential impact of VLAM is crucial.

Mapping an emerging field - the “Bell Labs 2.0” dynamic

Because VLAMs are still at an early stage, our analysis is deliberately exploratory. Rather than focusing on performance benchmarks, we map the emergence of the field along three dimensions: science, markets, and strategy.

First, we study the scientific trajectory of VLAMs using scientometric methods. Starting from a curated set of foundational papers and tracing more than 8,500 forward citations, we identify a rapidly expanding and increasingly competitive research landscape.

One of our most important findings concerns who produces frontier knowledge. In VLAMs, early breakthroughs overwhelmingly originate in corporate research labs, often in collaboration with universities, while public institutions dominate subsequent development and diffusion. We describe this pattern as a “Bell Labs 2.0” dynamic: a revival—under contemporary conditions—of corporate-led basic research. This challenges the conventional view that universities sit upstream of industry in the innovation pipeline and raises broader questions about the governance of scientific progress in AI-intensive domains. In fact, the VLAM case could be an early warning sign of a wider change in the regime of knowledge production in high-tech industries, characterised by industrialisation, narrowing and privatisation of fundamental research.

Who is riding the VLAM wave?

Second, we examine market formation around VLAM-powered robotics. Drawing on original data on startups, incumbents, partnerships, and funding, we find a fragmented but dynamic ecosystem. Importantly, much of the momentum comes from software-centric firms, including startups that develop foundation models or control layers for robots without producing hardware themselves.

This points to an ongoing softwarisation of robotics. While hardware remains essential, value creation increasingly concentrates in software, data, and models—mirroring developments observed earlier in sectors such as automotive and telecommunications. Notably, there is limited overlap between traditional robotics players, scientific contributors, and emerging VLAM firms, suggesting that VLAMs may not simply reinforce existing industry structures but instead introduce new market entry points as well as create novel incentives for industrial partnerships (as well as mergers and acquisitions, should incumbent firms in robotics decide to curb the risk of disruption).

Strategy, platforms, and gatekeeping

Third, we analyse strategy, focusing on how leading actors attempt to shape the VLAM ecosystem. Using Google DeepMind as a case study, we show how open datasets, open-source releases, and later proprietary reintegration coexist within a coherent strategic logic. Initiatives such as large shared datasets position external researchers as complementors; open-source models encourage adoption and standardisation; and proprietary platforms ultimately consolidate control.

We interpret these moves through the lens of platformisation strategy. By orchestrating an ecosystem around shared resources while retaining control over critical interfaces, a firm can accelerate innovation and simultaneously position itself as a gatekeeper. This dual dynamic—openness paired with control—has become a defining feature of competition in AI, and VLAMs appear to be no exception.

VLAMs as a systemic innovation for robotics

Our broader claim is that VLAMs are not just a technical advance, but potentially a systemic innovation with far-reaching implications for robotics. They illustrate how progress in upstream AI can transform downstream physical industries, how corporate labs can reclaim a central role in foundational research, and how platform strategies can reshape competition at an early stage of market formation or during the rejuvenation of a mature industry.

Whether VLAMs ultimately deliver on their promise remains an open question. But even at this early stage, they offer a revealing window into the future of robotics—and into the evolving relationship between science, markets, and strategy in the age of (embodied) AI.

References

When robots met language (models): An exploration of science and strategy in the vision-language-action models space PDF-download. 

*Quote from Vincent Vanhoucke’s talk “Robotics in the Age of Generative AI” at Nvidia GTC 2024

Co-authors

• Gabriela Fuentes, PhD Candidate in Economics, University Cote d'Azur and GREDEG/CNRS
• Taekyun Kim, Assistant Professor, Chungnam National University, School of Business