By Cristian Meo, CEO, LatentWorlds
The level of Physical Intelligence needed to deploy robots out of the box remains well beyond what current technology can achieve. Especially compared to the accelerated rate of technological innovation occurring within generative AI applications. I expect that this will begin to change in 2027.
I am seeing advances in physical AI, more robust and expanded training datasets, and improvements in robotics manufacturing that are bringing the industry closer to some exciting commercially viable deployments. However, success in this new era will not be defined by who develops the largest or most sophisticated model. Instead, it’ll be demonstrated by those who are translating intelligence into physical action within the practical constraints of real-world hardware.
Robotics Data Collection Will Accelerate, Not Plateau
In the physical intelligence world, there’s a big debate on whether the industry’s appetite for data will continue growing or reach a point of diminishing returns. I, personally, expect to see compelling evidence over the next few months that shows that we’re still at the beginning of the growth curve.
Unlike traditional LLMs, which learn from vast quantities of text and digital content, physical AI systems must understand how objects move, how environments change, and how actions produce consequences. Developing that understanding requires diverse datasets capturing interactions with the physical world, such as visual observations, movement trajectories, and sensor feedback.
Companies like 1X, Figure, SkildAI and FieldAI are investing in approaches to collecting and learning from physical-world data. Their efforts reflect proof that the industry has barely scratched the surface of the information needed to develop robots capable of operating reliably across different environments and tasks.
As more of these physical systems enter real-world settings, they’ll also create opportunities to collect additional data, improve models, and expand the range of tasks that the model itself can perform. This has the potential of generating a feedback loop in which deployment itself becomes an engine for further innovation.
I believe that 2027 will mark an important turning point. Physical AI models will become capable enough for a growing number of commercial applications. While improvements in production will make deployments economically feasible for a larger portion of the market. You can’t achieve general purpose robotics overnight. But I expect meaningful progress beyond controlled demos.
Smaller, Specialized Models Will Lead Real-World Deployments
Bigger is not necessarily better. This has been a tough reality for the AI industry to come to terms with as the race to build increasingly powerful foundational models has been widely demonstrated. But robotics is vastly different. Bigger models can introduce physical challenges that software developers don’t typically encounter.
Robots are unique. They need to process information and make decisions in real-time, often using onboard computing resources. Running a large model requires substantial processing power, which increases energy consumption, cooling requirements and hardware costs; these feed into a lot of the conversations the industry is elevating surrounding sustainability and cost efficiency. And it’s why I believe that the hardware constraints will become a defining factor in which physical AI technologies achieve commercial scale.
To set the record straight, large general-purpose models such as GPT-based systems and Google’s Astra genuinely demonstrate impressive capabilities. But the computational requirements make it difficult to deploy directly on today’s hardware. Distilling intelligence into smaller models that are truly optimized for specific environments can offer a more practical path. The winning approach will involve delivering sufficient intelligence while meeting the constraints of power, latency, cost, and physical design.
Training Smaller Models Will Become a Bigger Engineering Challenge
There is a misconception about small AI models that I want to address. They are not necessarily easier or cheaper to develop. Yes, they require fewer resources during deployment, but producing a compact model with strong capabilities can demand substantial investment in training infrastructure, data, and computation resources.
One emerging technique involves using larger models as essentially teachers to train smaller models. This approach empowers developers to transfer capabilities from computationally intensive systems into models designed to operate with tighter constraints. The challenge, however, is preserving performance and generalization. A smaller model that performs well on familiar tasks but struggles when conditions change offers limited value in unpredictable physical environments.
I expect more attention to engineering problems such as this. We need to be able to answer questions like “How can I compress sophisticated intelligence into efficient systems without sacrificing the adaptability robots need?”
At the end of the day, choosing the right model is an exercise in balancing capability against practicality. A larger model may offer stronger performance, but that advantage can come with hiccups. If the hardware required to run it makes deployment impractical, then the value is not realized.
In Summary: My 2027 Outlook
The next breakthrough in physical AI isn’t about becoming “smarter.” It will be defined by making sophisticated intelligence efficient, reliable, accessible, and affordable enough to operate in real environments! That’s where the robotics industry will begin turning years of research into tangible commercial progress.
##
ABOUT THE AUTHOR
CEO and Co-founder of LatentWorlds AI, the deployment backend for Physical AI. DataCore, their core product, handles robot fleet data infrastructure: robot policy CI/CD, edge-to-cloud capture, synchronized retrieval, and reproducible datasets for debugging and training, so teams can scale their fleets without scaling their operational burden. Cristian holds a PhD in World Modeling and Physical AI from TU Delft, where his research on generative world models, video prediction, and model-based decision making earned a Best Paper Award at the ICLR 2025 World Models Workshop. Prior to founding LatentWorlds, he spent time at Mila under Prof. Yoshua Bengio and Dr. Anirudh Goyal, and secured over €400,000 in competitive research grants including a major allocation on Snellius, the Dutch National Supercomputer.
