Physical AI Is Still Waiting for Its ChatGPT Moment

Physical AI has become one of the biggest bets in venture capital, with startups raising billions of dollars to bring the advances behind large language models into robots and other machines.

But while the investment boom is real, the technology still faces a major challenge: robots can move better than ever, but teaching them how to perform useful work reliably remains difficult.

That tension was on full display at last week’s Actuate conference, an event focused on developers building AI systems for robots. Organized by Foxglove, a company that helps physical AI developers manage and visualize data, the conference has grown rapidly since launching in 2023. This year, attendance reached about 1,500 people, roughly three times the size of the event’s first edition.

The excitement comes at a time when investors are pouring money into robotics. But recent market events also show how quickly expectations can change.

Unitree’s IPO Highlights the Robotics Challenge

China’s leading robot maker, Unitree, recently arrived on China’s equivalent of the Nasdaq and was valued at around $66 billion.

The excitement didn’t last long.

This week, Unitree lost nearly half of its value. Analysts pointed to a key problem facing the robotics industry: physical capabilities are improving quickly, but robots still struggle to perform valuable real-world tasks consistently.

In other words, building a robot that can walk, balance or move objects is one thing. Building one that can reliably complete a job for a customer is much harder.

Robotics Has a Data Problem

One of the biggest obstacles is training data.

AI systems need enormous amounts of high-quality data to become capable, but collecting useful data for physical environments is much more complicated than gathering text from the internet.

At Actuate, a booth from physical AI infrastructure company Avala summed up the issue with a simple message: “the robotics data crisis.”

The industry is still far from creating a general-purpose robot that can perform almost any task. Even attempts to train robots end-to-end for specific jobs have struggled to deliver reliable commercial products.

Developers are therefore looking to the lessons of frontier AI companies. That means building larger and more diverse datasets, experimenting with different training methods and developing better reinforcement-learning environments.

Harry Mellsop, founder of simulation startup Antioch, described the current stage of physical AI as its “GPT-2 era,” referring to the OpenAI model that came before ChatGPT.

His argument is that the industry still needs considerably more data and computing power before it reaches the next major stage. In particular, he pointed to GPUs optimized for ray tracing, which can help create highly detailed simulations for robots.

Self-Driving Cars May Have a Head Start

Autonomous vehicles are arguably further along than humanoid robots.

One reason is data. Cars driven by humans generate enormous amounts of information about real-world environments. Another is that autonomous driving primarily requires a vehicle to avoid collisions rather than physically manipulate objects.

Much of the software and infrastructure used to build physical AI systems has already emerged from the autonomous vehicle industry.

Foxglove is one example. The company was founded by former employees of Cruise, General Motors’ former self-driving vehicle operation.

Now, some autonomous vehicle companies believe their experience with machine learning could give them an advantage in humanoid robotics.

Tesla is already pursuing this strategy with its Optimus robot. Autonomous driving company Wayve and ride-sharing company Uber have also launched robotics labs focused on humanoid form factors as research and development projects.

Should AI or Hardware Come First?

Wayve CEO Alex Kendall believes the robotics industry should learn from autonomous vehicles.

He argued that manipulation robotics is comparable to self-driving technology several years ago. According to Kendall, some of the infrastructure—including data systems, simulation technology and machine-learning operations—could eventually be shared across different types of robots.

However, the models would still need specialized post-training depending on the physical form they control.

Kendall also believes it is too early to lock into a specific hardware platform. Sensors and other robotic components are developing rapidly, so a truly general AI model should ideally be capable of working across different hardware.

Not everyone agrees.

Hardware and AI May Need to Be Built Together

Théophile Gervet, CEO of humanoid robotics company Genesis AI, takes a different view.

Genesis AI raised a $105 million seed round this year, and Gervet believes the industry is still too early for a purely software-focused “brain strategy.”

Instead, he argues that companies have an opportunity to design hardware and AI systems together.

That debate reflects one of the biggest questions in physical AI: should startups build a general-purpose intelligence first and then put it into different robots, or should they design the robot and its AI as one integrated system?

Narrow Robots Are Already Doing Real Work

While general-purpose humanoids remain largely in research environments, robots designed for specific jobs are already being deployed.

For example:

• Gritt is building robots for solar farms.
• Agility is deploying robots in industrial environments.
• Bedrock is operating excavators autonomously.

Gervet believes this creates an important dilemma for robotics startups.

Customers may not care about a robot that can perform dozens of different jobs if it only succeeds 80% of the time. A specialized robot that reliably completes one valuable task could be much more useful.

At the same time, focusing too narrowly could create another problem. A company that builds a vertical-specific robot on an older AI model could eventually be overtaken by a competitor using a much more capable general-purpose model.

Real-World Deployments Could Create the Data Advantage

There is a strong reason for robotics companies to focus on specific industries: deployment generates data.

A robot working on solar farms, in construction or inside a warehouse can collect information from real-world environments that is difficult to reproduce in a laboratory.

Bedrock CTO Kevin Peterson said the company initially focused on excavation to understand the challenges of “manipulation in the wild.” The longer-term goal, however, is to create an intelligence layer that can work across several types of construction machines.

The challenge is managing all that information.

Robots can generate huge amounts of visual and lidar data, making it difficult for engineers to find useful information and understand what went wrong.

Better Data Tools Could Speed Up Robotics

Foxglove is working on that problem as well.

The company announced a new product built on Nvidia’s Cosmos open-weight world model. The tool allows engineers to search large robotics datasets using sophisticated natural-language queries.

The aim is to make it easier to create evaluations and simulations, while speeding up the process of finding problems and debugging AI models.

For robotics companies, faster data analysis could mean faster development cycles and more frequent improvements.

What Will Be Robotics’ ChatGPT Moment?

OpenAI CEO Sam Altman has suggested that a major breakthrough in physical AI could arrive within a few years. But exactly what that breakthrough will look like remains an open question.

Kendall noted that the world’s largest robot deployments are still consumer robot vacuum cleaners.

For him, the real breakthrough will be something that excites ordinary consumers—not just investors, who are already highly enthusiastic about robotics.

One example he offered is achieving eyes-off autonomy for less than $1,000 worth of hardware in a car. Wayve is working toward that goal and licenses its models to automakers. Kendall sees the opportunity as potentially worth billions of dollars and believes it could help the company develop a truly general embodied AI model.

For Gervet, the defining moment would look different.

He imagines a robot that can understand natural-language instructions and reliably perform basic physical tasks such as pushing, pulling, closing a laptop or cleaning a table. If such a system could complete those tasks with around 80% or better reliability straight out of the box, he says that would feel similar to the ChatGPT experience.

Foxglove CEO Adrian Macneil has a different vision altogether.

He doesn’t believe robotics will necessarily have a single ChatGPT-style moment. ChatGPT’s impact was partly driven by distribution: it reportedly reached around one million active users in roughly a week.

Getting robots into the physical world is much harder.

Instead, Macneil is looking for what he calls an “Apple II moment” or “IBM PC moment” for robotics—the point when consumers can simply buy a useful home robot that does interesting and practical things.

For now, the robotics industry is still searching for that moment. Investors are already convinced the opportunity is enormous. The harder task is proving that today’s physical AI can move from impressive demonstrations to robots that consistently deliver real-world value.

Also read : General Motors EV Brake Problems Face Tougher Federal Investigation