Physical AI's Second Wave Isn't Just About the Robots

By Navya Prabhushankar

Our focus at Interplay has always been software, and increasingly, application-layer AI. It’s the layer sitting closest to the end user, translating raw model capability into something that actually gets adopted. Physical AI looks like it belongs to a different category of investor, all chassis and actuators and factory floors. But not really. Hardware is commoditizing fast. Humanoid robot selling prices are expected to fall from $114,700 in 2024 to approximately $37,000 by 2030. Physical AI is a software-enabled hardware category, and the differentiation is migrating up the stack the same way it did in cloud infrastructure and fintech before it.

Robotics and physical AI startups raised a record $27.6 billion in 2025, more than double 2024, and blew past that entire annual total in the first half of 2026 alone. Almost all of it went to one question: can a robot perceive and act in the world at all? That question is closing. The harder one is what happens after a robot already works in a demo and has to survive contact with a real operation. Robots don't usually fail because they don't work. Poor system integration and mismatched ROI expectations alone account for more than half of robot deployment failures, with inadequate change management responsible for roughly a fifth on top of that. The technology almost never fails. The deployment does. That gap is a software problem, not a hardware one, and it's where we're looking.

Wave One

A useful way to think about physical AI is as two jobs stacked on top of each other. The first is perception, giving a machine the ability to notice what's happening around it. The second is action, giving it the ability to do something about it. Nearly all of the capital and the noise has gone into the first job and the infrastructure underneath it: foundation models, world models, sensors, chips.

That job is also dwindling, at least the horizontal part of it. Vision transformers have mostly replaced the older, narrower detection models, and open vocabulary systems can now find almost anything described in plain language without being trained on it first. A custom inspection model that needed 50,000 labeled images in 2023 can be fine-tuned with 2,000 to 5,000 today, cutting data collection costs by roughly 90 percent. Hundreds of funded companies, 463 by one recent count, are building on the same handful of foundation models. Generic camera-based perception is not a differentiator anymore.

Getting a robot to work is wave one. Getting it to stay working, inside a real operation, with real people, is a different problem entirely, and it's the one that needs to be solved.

What’s Breaking

Robots break on multiple levels. Battery life, hardware durability, and compute are real limits. So is the sim-to-real gap, though that space is already crowded, as NVIDIA's simulation platforms have largely claimed it. We're focused on what happens after all of that already works, once a robot is good enough to leave the lab. That's not a small population anymore. Roughly 542,000 industrial robots were installed worldwide in 2024 alone, pushing the global operational stock past 4.6 million units. Every one of them entered exactly this integration gap the moment it went from pilot to production, whether or not it made it through.

A pilot usually gets approved running one robot on one task, coordinated by the same people who designed it specifically to succeed. Production means a dozen robots from two or three vendors, none of which speak the same protocol, yet all of which need to plug into a warehouse management or ERP system that was never built with a robot in mind. That integration cost is rarely in the original business case, so the ROI model that got the pilot approved quietly stops matching reality the moment it has to scale. Deployment timelines commonly stretch from six weeks to six months as a result, not because a robot broke, but because nobody priced in the cost of making it fit.

Where We’re Looking in Wave Two

Orchestration and integration infrastructure. This is the software layer that makes a robot deployment behave like one coordinated system instead of a collection of point products. It connects fleets to the enterprise systems already running a facility, and makes it fast and cheap to redeploy a robot for a new task instead of custom-coding it each time. InOrbit handles the first half, sitting above individual vendors' fleet software and translating orders from warehouse, ERP, and manufacturing execution systems into coordinated action across more than 30 robot brands, for customers including Colgate-Palmolive and Genentech. Waddle Labs is tackling the second, using coding agents to write and test a robot's control behavior for a new task directly on real hardware, cutting a process that's traditionally been slow and brittle down to minutes. Different layer, same underlying cost, and it's exactly the cost that erodes the ROI that automation was supposed to deliver.

Trust, override, and safety infrastructure. The clearest slice of the change-management cause. Warehouses adding robots saw severe injuries drop 40 percent, while non-severe injuries rose 77 percent, concentrated during peak periods when pick rates ran two to three times higher than normal. Research on collaborative robots finds resistance is rooted less in fear of robots generally than in workers not understanding what a machine is doing. A separate study found a manual override was what workers said they needed to trust one at all. Wireless e-stops aren't new. Companies like Cattron have been selling SIL-3-rated versions for years. What's new is treating this as software-defined infrastructure for fleets rather than a hardware accessory for one operator, which is what FORT Robotics has built, with roughly 12,000 units deployed across 600+ customers since 2018.

The regulatory ground is shifting too, the EU AI Act classifies autonomous robots as high-risk with obligations phasing in from 2027, and ISO's core human-robot collaboration standard just underwent its first major revision in years. Saphira builds a risk engineering layer for autonomous machines, turning system design into living risk models and generating the evidence regulators and insurers demand, with humanoid and industrial robotics customers including 1X Technologies, RobCo, and Reflex Robotics already live.

First contact. A different kind of trust than the above. The other categories are about coexistence; can a person near a robot read it, stop it, trust it in the moment. This one is about delegation - whether a professional or a trade is willing to hand over work that used to require a skilled crew's judgment, with no one standing there to verify it went right until after the fact. Terranova is asking a city engineer to trust autonomous injection robots with land-raising work that previously took a century of manual civil engineering. Salem Robotics is asking nuclear facilities to trust autonomous systems with radiological surveys and compliance recordkeeping. Such work has historically been done by a human walking a survey route with a handheld instrument, precisely because the cost of getting it wrong has always been too high. Nothing about the underlying shift is new, cheaper hardware, more capable software. What's new is the trade, and the trust it's asking for is harder to earn than anything else here.

What Makes Us Move

1. Works across robot platforms. Not bundled into one company's hardware, the value of a shared layer disappears if it only works with one machine.
2. Proprietary data that compounds with deployment. Human-interaction data, sensing data no foundation model has already absorbed, and threat data from real fleet attacks. The moat is the accumulation, not the algorithm.
3. Evidence tied to deployment outcomes. Fewer canceled pilots, higher renewal rates, and lower incident counts - not a user study run in a lab.
4. Real path into existing distribution. This layer only matters once it's sitting on top of a machine that's actually out in the world, so a way into robotics companies with deployment already underway matters more than a standalone go-to-market plan.
5. Trade credibility. A founding team with real standing in the industry being automated, plus a committed customer or LOI in hand, not just a demo.

Physical AI has really been asking one question in two different forms. The first form was whether a machine could work at all. The second is whether it can survive the mess of an actual operation, the integration, the trust, and in a few remaining trades, whether to let one in for the first time.

If you're building any part of that answer, we'd like to hear from you. Reach out to our team at np@interplay.vc.