Demo to Deployment: Paras Savnani on the Testing Gap in Robotics

The robotics founder argues that impressive demonstrations reveal far less than the industry needs to know about whether a robot can perform reliably once it leaves a controlled environment.

A robot can complete a polished demonstration and still be far from ready for everyday use. Paras Savnani has spent nearly a decade working across robotics and artificial intelligence, and he sees that gap as a major obstacle to wider robot deployment. Now the founder and CEO of Robolens (Lunaris AI), Savnani is building testing infrastructure intended to shorten a development process that still depends heavily on slow real-world evaluation. His concern is straightforward: robotics companies are getting better at showing what machines can do, but demonstrating a capability once is very different from proving that it will work reliably after deployment.

“A good demo proves that the robot can do something under those conditions,” Savnani said. “It does not tell you how often it will fail, what happens when conditions change, or whether you can trust the behavior repeatedly in production.” He believes much of the industry conversation still concentrates on visible breakthroughs while giving less attention to the infrastructure required to evaluate robot policies repeatedly and learn systematically from failures.

The challenge becomes more difficult because robotics development still depends heavily on real-world testing. Savnani argues that this requirement makes development and deployment cycles slow, particularly when teams need to evaluate new models or policies repeatedly. His current work is based on the premise that a larger share of evaluation could happen in simulation, with last-mile testing remaining in the real world.

“Robotics cannot iterate at software speed if every meaningful test has to happen physically,” he said. “The question is how much of that evaluation you can move into a controlled digital environment without losing the information you actually need.” Savnani does not argue that simulation can simply replace reality. He sees a nearer-term opportunity to use it for evaluation, allowing developers to explore many more scenarios before returning to physical systems for last-mile testing.

Savnani’s interest in that problem is informed by experience on both sides of physical AI. He has spent almost 10 years working across autonomous robotics, machine-learning research, and deployed AI systems, including graduate work in robotics at the University of Maryland. That range has given him experience with both the models that drive intelligent behavior and the physical systems that ultimately have to perform reliably.

Robotics teams may have sophisticated models, but he says there is still no standard way to evaluate frontier systems across different robot embodiments. Companies frequently assemble their own in-house pipelines, and different robot embodiments and environments make a single industry-wide benchmark difficult to establish.

“Two teams can both say their robot works, but they may be measuring completely different things,” Savnani said. "Every robot is different, and the evaluation has to reflect what that robot will actually be expected to do." His view is that testing should begin early enough to influence development rather than appear as a final hurdle immediately before deployment.

That principle also shapes his advice to robotics teams. Savnani favors getting robots into the field quickly enough to learn from real production conditions, especially in non-critical environments where early failures can be tolerated and studied. At the same time, he believes many scenarios can be reproduced through simulation or in-house replicas, giving teams more opportunities to test behavior without waiting for every failure to occur naturally.

He also rejects the idea that robotics deployment should be viewed as a binary choice between full autonomy and human control. Savnani sees autonomy as a sliding scale, particularly while general-purpose robots are still improving. Early systems may require people to monitor fleets, intervene when machines become stuck, and support recovery after failures. “A robot does not need to be perfectly autonomous on day one to be useful,” he said. “What matters is understanding where it can operate safely, where it is likely to fail, and what recovery looks like when that happens.”

Robolens is roughly three and a half months into development, with a functioning minimum viable product and an active effort to work with design partners to test whether the platform creates value inside existing robotics workflows. The next step is to validate that usefulness before attempting broader scale through a self-serve product.

Savnani’s larger argument is that robotics will not reach widespread deployment through better demonstrations alone. Teams need ways to evaluate failure, compare behavior across changing conditions, and feed what happens in production back into development. He believes simulation can become an important part of that loop, particularly as companies search for faster ways to test increasingly capable systems.

“The exciting part of robotics is seeing what a machine can do,” Savnani said. “The work that determines whether it reaches the real world is proving that it can keep doing it when the conditions stop being perfect.” For him, that is where the next major advance may be required: building the testing infrastructure that makes reliable deployment possible.