Physical AI's Biggest Problem? Teaching Machines How to Feel

Physical AI is driving record-breaking surge in robotics funding as investors pour money into AI models that let machines generalise across hardware rather than building expertise in single-task industrial arms. 

Crunchbase pegged global robotics funding at $18.8 billion for 2026 through late June, already ahead of all of 2025's $15 billion. Meanwhile, Goldman Sachs has projected cumulative investment in humanoid robotics alone will surpass $50 billion by 2030. 

Capital is pouring into both embodied foundational models as well as edge infrastructure.

Travis Kalanick's industrial robotics venture Atoms—which aims to automate sectors like food production, mining, and transportation—closed a $1.7 billion round led by Andreessen Horowitz in July. Meanwhile, Skild AI tripled its valuation in just seven months after raising $1.4 billion in January to create an “omni-bodied” foundational model that works across all physical AI hardware.

However, almost none of that money is fixing the data gap that serious players are grappling with. Unlike text or images, physical interaction data cannot be scraped from the internet. Every example has to be performed, recorded, and annotated by hand. Data has become the defining constraint in physical AI even as capital floods in.

Deccan AI, a Bay Area-headquartered company with its largest operations base in Hyderabad, is one of a growing number of firms trying to solve that constraint. It closed a $25 million Series A round in March, led by A91 Partners, to scale post-training data operations and expand into robotics data. 

The company builds post-training data, evaluations, and reinforcement learning environments for AI labs and enterprises, drawing on a contributor network of over a million people. Its reported client roster includes Google DeepMind and Snowflake.

Scaling AI Doesn’t Work for Robots

According to Mordor Intelligence, the data annotation market crossed the $3 billion mark globally in 2026, with computer vision dominating the bulk of annotation tasks.

Navaneeth Bodla, Head of AI at Deccan AI and who earlier built models at self-driving vehicle company Cruise and Adobe, describes physical AI being at something like a GPT-2 moment. There’s early evidence that scaling, in the form of bigger models and more data, works for embodied intelligence the way it worked for language. 

But there’s a catch. 

"You also need to have inverse scaling loss, where your hardware, compute are also catching up," he tells AIM, pointing to the need to shrink compute and hardware so that these larger models can actually run on a physical robot without picking up subtle noise that overcomplicates its motions.

Solving that dual requirement—bigger models paired with smaller, deployable hardware—is part of the reason why the industry has leaned so heavily on humanoid form factors over the past two years.

The Data Mix

Deccan AI produced the multimodal training data for companies to ensure that the robotics hardware has real-world understanding. That spans egocentric video captured from head- or chest-mounted cameras, multi-camera rigs that add depth information, and motion capture setups using 23 cameras plus inertial sensors for millimetre-level tracking of body movement. It collects the data both in-house and with partners.

Nikhil Parlapalli, who leads the compute and deployment side of Deccan AI's robotics work, says the startup effectively serves two kinds of customers. One set is frontier AI labs trying to extend their models into robotics. The other is robotics native companies training their own smaller, purpose-built models. 

"Both of them need training data sets which are not easy to scale," he adds. "You need to actually go out and collect this physical data, which is a difficult process." 

Beyond Deccan AI’s competitors, many other startups are also dabbling in this field, although at the risk of data privacy. 

Quick service startup Pronto runs a limited physical AI data pilot programme after raising $20 million for physical AI-focused investor Lachy Groom. Pronto earlier told AIM that it had a “strictly opt-in” pilot involving outward-facing wearable cameras used by professionals during specific jobs.  

Limits of Embodied Data Collection

The hardest data to collect, according to Bodla, is not visual at all. It is touch. Cameras, however many angles they cover, can record what a hand does. They cannot record what it feels.

"When you pick up a phone and put it down, you feel the pressure in your hand, that's information you will not be able to collect purely by collecting an egocentric video of it," he said. Weight, temperature, and whether an object is soft enough to squish cannot be determined just from video. 

While sensors can capture information covering pressure, temperature, and force feedback, they are not yet easy to scale or integrate into natural hand movement—a problem multiple vendors in this space describe as the current frontier of embodied data collection.

Deccan AI's workaround is to strip a robot down to its most essential components, typically the hands and the most critical cameras, and use teleoperation to collect data through that reduced rig instead of instrumenting every sensor a full robot might eventually carry. 

Parlapalli, who earlier led camera application development for autonomous vehicle systems at Magna International, frames the stakes of getting this wrong in physical rather than abstract terms. In robotics, he says, the risk that matters is of a model failing at an inopportune moment while interacting with humans. 

"It's not that it is going wrong, it's just failing at the wrong point," he notes, comparing it to guardrail engineering in autonomous vehicles, where the priority has always been preventing a system from causing harm, not preventing some hypothetical takeover scenario. Better tactile and force data, in that framing, is not just a capability problem. It is a safety one.

An Early, Expensive Field

Gopi Pepakayala, who joined Deccan AI after over six years at AWS working with VC-backed startups, is cautious about how this settles. "We don't know what ultimate architecture, or what type of model, or what type of data set it's finally going to result in," he shares. "It's a very early stage right now."

That caution sits oddly alongside the money moving into the sector. Physical AI's funding curve has outpaced its data infrastructure by a wide margin. Scale AI and Appen dominate the managed end of the AI data market. Companies like Shaip and Encord are building out annotation and orchestration tooling, and specialist collectors are pitching everything from teleoperation programmes to human video sourcing. 

Deccan AI is competing for a slice of that pipeline from an India-based operations base, at a moment when, by Pepakayala's own account, nobody, including Deccan, actually knows which architecture or data mix wins.

"Deccan wants to make sure that we are in the right place in this domain," he affirms.