The embodied AI sector, which connects machine-learning models to mobile bodies and robots, is experiencing a strong investment wave, but advances in hardware have not yet been matched by comparable progress in reliably completing valuable work. This gap summarizes the sector’s current position: robots have become more physically capable, but giving them the knowledge required to deal with the real world remains an open challenge.
This contradiction was clear at the recent Actuate conference, organized by Foxglove for developers of models and infrastructure for embodied AI. The conference has tripled in size since its launch in 2023 and attracted 1,500 participants, signaling growing interest in the field. At the same time, infrastructure companies focused on what they describe as the robotics data crisis: the shortage of the high-quality, diverse training data that models need.
Why Aren’t Physical Capabilities Enough?
Attempts to build a general-purpose robot capable of carrying out any task remain far from mature. Even when a robot is trained on a specific task using end-to-end learning, these approaches have not yet produced products with consistent, dependable commercial performance. Developers are therefore turning toward approaches that echo what advanced AI laboratories have done: collecting or creating more diverse datasets, testing multiple training methods, and designing better scenarios for reinforcement learning.
Harry Mellsop, founder of Antioch, a company specializing in simulation tools for model developers, described embodied AI as being in its “GPT-2” stage—the period before the leap that made ChatGPT a mass-market model. Under this view, the sector needs more data and computing power to move beyond its current stage, including graphics processing units optimized for the ray-tracing techniques used to build high-fidelity simulations.
But comparing the field with language models does not mean the path will be identical. A robot interacts with a changing physical world and needs visual data, LiDAR data, and behaviors connected to the environment—not just text. A failure in a physical task can also lead to a collision, damage, or an operational shutdown, making the transition from a model that works occasionally to a reliable commercial product more complicated.
Autonomous Vehicles as an Early Laboratory
Autonomous vehicles appear to be among the most advanced areas of embodied AI, partly because they can collect large quantities of data from cars driven by people, and because their core task centers on avoiding collisions rather than directly handling and moving objects. Many tools for building models have moved into robotics from autonomous-driving companies; Foxglove, for example, was founded by former employees of Cruise, General Motors’ former self-driving project.
Automotive companies are now trying to use this expertise to compete in humanoid robots. Tesla is developing the Optimus robot, while Wayve, an autonomous-driving company, and Uber have launched robotics laboratories focused on humanoid bodies as research and development efforts.
Alex Kendall, Wayve’s chief executive, believes that starting with vehicles may make more sense because the data infrastructure, simulation, and machine-learning operations can be shared in part across applications. But he emphasizes that the universal model will require differences tied to each body, robot, and operating environment, and that rapid advances in sensors and components make committing early to a single hardware platform a risky decision.
Specialization Versus the General-Purpose Robot
Not everyone in the sector agrees with the idea of separating the “mind” from the body. Théophile Gervet, chief executive of Genesis AI, a vertically integrated humanoid-robotics company, believes the wave is still at a sufficiently early stage to allow hardware and AI to be designed together. The company raised a $105 million seed funding round this year.
The disagreement is not only about engineering, but also about business-building strategy. Companies targeting specific tasks have begun moving their robots into real-world locations: Gritt is working on solar farms, Agility is deploying robots in industrial environments, while Bedrock operates autonomous excavators. By contrast, multipurpose humanoid robots remain largely inside laboratories.
Gervet says customers do not care about a general-purpose robot that succeeds at a task only 80% of the time, because a lack of sector focus may mean a lack of practical value. Specialization, however, creates another problem: if a vertical product is built on top of an early generation of models, a competitor using a more advanced model may surpass it. Even so, a specialized application provides revenue and real-world data, even if that data is not diverse enough to develop a general-purpose robot.
Bedrock is starting with excavation to understand the challenges of handling objects in real conditions, but it plans to build an intelligence layer that extends across a range of construction machines. This illustrates a possible compromise: launch a specialized product that can gather experience and data while retaining the goal of building broader capabilities later.
What Is Changing in Practice?
Infrastructure is moving toward addressing the data problem itself. Foxglove announced a new product built on top of Nvidia’s open-weights Cosmos world model, allowing engineers to search vision and LiDAR data using advanced natural-language queries, with the aim of creating evaluations and simulations and speeding up the sorting and correction of problems. The value of these tools lies not in adding physical capability to the robot, but in reducing the time needed to identify why a model failed and retrain it.
The broader question is what will create the robotics equivalent of the ChatGPT moment. Kendall believes the real moment must attract consumers, not investors, and offers as an example unsupervised self-driving without eyes-on monitoring using hardware costing less than $1,000—an opportunity Wayve says it is pursuing by licensing its models to automakers. Gervet believes the defining milestone will be a robot that understands natural commands and performs basic manipulation tasks, such as pushing and pulling, closing a laptop, or cleaning a table, with a success rate of around 80% or more without complicated setup.
But Adrian Macneil, Foxglove’s chief executive, does not expect a single moment comparable to ChatGPT; deployment in the physical world is much harder than distributing software online. Instead, he expects a moment resembling the emergence of the Apple II or IBM PC, when it becomes possible to buy a household robot that performs useful and enjoyable tasks.
The most important reading of these positions is that the sector has not yet settled whether progress will come from a general-purpose model, co-design between hardware and software, or specialized applications that expand gradually. What the available evidence confirms is that increased investment and improved bodies do not eliminate the core bottleneck: realistic and diverse data, accurate simulation, and performance that customers can rely on outside the laboratory.