The diversity of robot training data is the single most important factor for achieving generalizability, surpassing the importance of sheer data quantity.
The primary risk in building general physical intelligence is the immense technical difficulty of the problem itself, not market competition.
An open strategy of sharing research, models, and even hardware designs is the most effective way to attract the top-tier talent necessary to solve general-purpose robotics.
Robots require data from their own physical embodiment to learn effectively; data from observing humans is valuable but insufficient for mastering motor control.
Humanoid robots are currently 'overrated' due to the difficulty of teleoperation for data collection, and the future of robotics will be a 'Cambrian explosion' of diverse, specialized hardware forms.
Pre-Physical Intelligence
Fent cites foundational research projects like Aloha, Mobile Aloha, and RTX, which demonstrated the feasibility of using teleoperation for data collection and the benefits of pooling data across different robot embodiments.
Experience at Large Companies
Fent describes her experience at large companies like Google, where code security policies made it 'nearly impossible' to take robots off-campus, severely hindering the ability to gather the diverse data she believes is essential for progress.
Founding of Physical Intelligence
Fent co-founds Physical Intelligence with the explicit goal of building a general-purpose AI model for all robots, contrasting with the industry's historical focus on specialized, single-application solutions.
Current Strategy & Operations
The company's primary method for data collection is teleoperation of real, often inexpensive, robots. They have adopted an open strategy, publishing research and open-sourcing models to attract talent.
Immediate Technical Priority
Fent identifies adding memory to their current memoryless, transformer-based policies as a higher priority than integrating new sensor modalities like tactile sensors, which she views as not yet commercially viable.
Future Outlook
Fent references Sergey Levine's prediction that the achievement of general-purpose robot intelligence will lead to a 'Cambrian explosion' of diverse, specialized robot hardware, moving beyond the current focus on humanoids.
▶Data Diversity as the Cornerstone of GeneralizationMar 2026
Fent consistently argues that the most critical factor for creating general-purpose robots is the diversity of training data, not just the volume. This includes data from a wide variety of robot platforms and physical scenarios, as demonstrated by the success of the RTX project which pooled data from different embodiments.
Investors should scrutinize a robotics company's data acquisition strategy, prioritizing those with a clear, scalable plan for capturing diverse, embodied data over those simply accumulating large quantities of homogenous data.
▶The Primacy of Embodied ExperienceMar 2026
A core tenet of Fent's philosophy is that robots must learn from their own physical interactions with the world. While observational data from humans is useful, she asserts it is not a substitute for the data a robot generates from its own body, which is essential for learning motor control.
This focus on embodied data suggests that companies with robust physical infrastructure for teleoperation and real-world data collection may have a long-term advantage over those relying primarily on simulation or observational datasets.
▶Openness as a Strategic ImperativeMar 2026
Fent's company, Physical Intelligence, intentionally pursues an open strategy, including publishing papers, open-sourcing model weights, and sharing robot designs. This is framed not as a purely academic exercise, but as a crucial business strategy to attract and retain the elite talent required to solve such a difficult technical problem.
This 'open' model challenges traditional IP-hoarding strategies, suggesting that in deep-tech fields, talent acquisition and ecosystem building can be more valuable assets than proprietary code or models.
▶Pragmatic Hardware and Model DevelopmentMar 2026
Fent expresses skepticism about the current hype around humanoid robots and the viability of complex sensors, viewing them as impractical for data collection. Instead, her approach prioritizes cheap, simple robots for rapid and diverse data gathering, and focuses on adding foundational capabilities like memory to models before incorporating new sensor modalities.
This pragmatic approach signals a focus on near-term data acquisition velocity over pursuing the most advanced hardware, which could lead to faster model improvement cycles and a more capital-efficient path to general intelligence.