Collecting more robotics data does not necessarily produce a better model. Paras Savnani has spent much of the past decade working across robotics, machine learning, and generative AI, and he argues that the quality and composition of training data can matter as much as sheer volume. That question sits alongside another technical debate he is following closely: how physical-AI systems should represent the world they are expected to operate in.
“More data is not automatically the same as better data,” Savnani said. “You have to ask what the robot is actually learning from and whether that information represents the situations you want it to handle.” He pushes back on the idea that volume alone determines the value of robotics data. In his view, carefully curated, high-quality data can produce stronger results than indiscriminately accumulating more examples. The marginal value of data often comes from novelty: redundant examples add less, while rare states, failures, and underrepresented behaviors can carry disproportionate signal.
Different companies are placing different bets on how robotics data should be collected and used. Physical Intelligence has emphasized heterogeneous data across tasks and robot embodiments, Sunday Robotics uses sensorized human demonstrations through its Skill Capture Glove, and Dyna-2 was pretrained on more than one million hours of egocentric human video. For Savnani, these approaches reinforce the idea that data strategy is not simply a race for scale: what is collected, how diverse it is, and which failures or rare states it captures can matter as much as raw volume.
“The important question is not only how much you collect,” he said. “It is what kind of experience the data gives the model.” Savnani’s view is that robotics teams need to think carefully about the type and quality of information they collect rather than treating total volume as the primary measure of a useful dataset. The different approaches being explored across the industry reinforce his view that data strategy remains an open technical question.
World models create another area of disagreement. Savnani is particularly interested in the split between approaches that generate future observations, such as video, and approaches that predict in latent representation space, such as JEPA. He currently favors the latter for planning and is watching how both directions develop as physical AI advances.
“There are different bets being made on world models right now,” Savnani said. “I am more interested in the latent-model direction, and I think it is important to see how these approaches develop as physical AI advances.”
His position is informed partly by having worked directly with one side of that technical landscape earlier in his career. At Samsung Research’s NEON Labs, Savnani trained foundational motion and video-generation models for real-time AI avatars that were deployed for enterprise users. That work gave him direct experience developing motion and video-generation systems.
His broader interest lies in how those technical choices connect with the other unresolved pieces of robotics development. Savnani sees data and evaluation as two major gaps that still need to be addressed if general-purpose robots are to become more common. He believes both are central to helping robotics teams iterate more quickly and develop more capable physical-AI systems.
Simulation is another area where Savnani sees unanswered questions. He considers the sim-to-real gap an important problem and believes simulation may have a particularly useful near-term role in evaluation, even as companies continue exploring it for training. His interest is in understanding where simulated testing provides useful information and where real-world validation remains necessary.
That uncertainty is one reason Savnani remains cautious about treating any one technical approach as settled. Physical AI is developing through parallel experiments in data collection, model design, simulation, and evaluation. His preference is to examine those choices through what they enable rather than assume that scale by itself will resolve the underlying technical questions.
Savnani estimates that broadly useful general-purpose robots could reach meaningful deployment within roughly five years. He points to 1X, already testing its NEO robot in real homes ahead of planned consumer deliveries in 2026, as a sign that robots are moving beyond industrial settings. At the same time, he sees unresolved questions around the data, evaluation methods, and technical approaches that will support broader deployment.
“I do think general-purpose robots are coming,” Savnani said. “The interesting question is what combination of data, models, and evaluation gets us there.” His forecast reflects optimism about the direction of the field while leaving room for the possibility that current strategies will continue changing as teams learn more about what works.
Savnani’s own technical record gives him experience across several parts of that debate. He began with autonomous robotics, later worked in machine-learning research, and trained generative motion and video models before deploying AI systems in other settings. That range has made him interested in the relationships between these technical choices rather than treating any one of them as sufficient on its own.
For Savnani, the unanswered questions run deeper than dataset size alone. He is watching how teams choose their data, which world-model approaches prove useful, and how simulation fits alongside real-world experience. Data and evaluation remain the two areas he sees as especially important to the next stage of physical AI, and he believes progress in both will shape the path toward more capable general-purpose robots.
