The humanoid robot race is becoming a race for data

Kara Quast

3 min read

The factory worker of the future is being trained in Georgia. It is six feet tall, powered by AI and called Atlas.

Boston Dynamics, a subsidiary of Hyundai, has opened a training centre at a Hyundai manufacturing site to prepare its general-purpose humanoid robot for industrial work. Here, Atlas is not yet being used for production. Instead the training centre is dedicated to the essential task of generating the data needed to train the AI models that will control Atlas.

Physical AI faces a fundamental data problem. Large language models such as those behind ChatGPT and Claude benefited from an enormous pre-existing training resource: an internet containing trillions of words and images.  Robots have no comparable store of physical experience. Their training data must generally be deliberately created, recorded and processed.

There is no equivalent to the internet providing data on every way to grasp a component, move through a crowded workspace or recover when an object slips. Building robot hardware is therefore only part of the challenge. AI models must also be trained to control that hardware safely and reliably across thousands of physical tasks.
Before companies can deploy general-purpose robots at scale, they may first need to build a larger database of physical data.

Hyundai and Boston Dynamics are trying to address that problem in Georgia. Their approach illustrates why the next phase of the robotics race may be determined less by who builds the most capable hardware and more by who can generate the best data to control it.

Why physical AI data is hard to produce

Unlike language models, robots need data that connects what they perceive with movements, forces and physical outcomes. This can include camera images, joint movements, sensory feedback, human demonstrations, teleoperation and simulated or real task attempts.

Producing this data is slow and expensive. Real-world experience requires physical robots and often human operators, while simulation can generate data at scale but may not transfer perfectly to the real world.

Sharing data is also difficult because researchers use different robots, sensors and data formats. The challenge, then, is not simply producing more data, but producing enough diverse, compatible and physically meaningful data.

Hyundai and Boston Dynamics: turning factories into data sources

This is where Hyundai Motor Group may have a structural advantage: it owns the type of environment in which its robots are ultimately intended to operate.

Boston Dynamics is using Atlas robots at a dedicated training centre in Georgia that replicates key elements of Hyundai’s manufacturing environment. The purpose is not primarily to produce cars, but to generate data for training the AI models used to control Atlas.

The robots perform manufacturing tasks while data is collected about their movements, interactions with objects and the outcomes of those actions. This can be combined with data from human demonstrations, teleoperation and simulation to train and improve the models that determine how Atlas performs physical tasks.

Access to manufacturing infrastructure is particularly valuable here. Hyundai can provide realistic components, processes and factory conditions in which useful training data can be generated. Once robots are deployed in actual factories, real manufacturing operations will provide an additional source of data for further model training.

Atlas robots are intended to begin parts-sequencing work in 2028, expanding into component assembly from 2030. Hyundai intends to establish annual production capacity of 30,000 robot units by 2028.

The long-term advantage may therefore belong to companies that have access to environments where robots can be used to generate large quantities of useful, real-world training data.

Different approaches to the physical AI data problem

Hyundai and Boston Dynamics are not alone in trying to solve the shortage of physical-AI training data, though others are approaching the problem in different ways.

One such approach is simulation. NVIDIA, for example, has released an open dataset containing 24,000 humanoid motion trajectories and developed GR00T-Dreams to generate synthetic motion data. Simulation can produce large quantities of training data without requiring the equivalent amount of time on physical robots.

Another approach is to capture what video misses from human demonstrations. A camera can show a hand picking up an object, but not how firmly it grips it, and the hand may disappear from view inside a box or behind another object. That missing information matters because apparently similar movements can require very different forces, whether turning a stiff control, handling fragile material or inserting a tightly fitting component.

Axonnic is exploring whether muscle signals can fill some of these gaps. Its forearm sleeve measures muscle activity to track finger movement and estimate fingertip force, synchronising this information with video and hand-pose data. The aim is to create richer data about how humans perform physical tasks. The company also says the signals can be used to control a robotic hand in real time.

Axonnic is still at an early stage, but its approach offers an exciting new way to address the physical AI data challenge. If successful, richer data on muscle activity and force could help capture aspects of human dexterity that conventional video misses.

Together, these approaches illustrate the broader challenge: there may be no single source of training data for physical AI. Real-world robot operation, simulation and richer recordings of human demonstrations may each provide different pieces of the data needed to train models that can reliably control generalised robots in the physical world.

Conclusion

Despite these promising developments, important challenges remain. Models trained on highly structured automotive tasks may struggle to generalise to the more unpredictable environments found in warehouses, construction sites or homes. At the same time, differences in sensors, coordinate systems and annotations make it difficult to combine and reuse training data across the industry.

Humanoid robotics is often presented as a hardware contest: stronger actuators, better hands, longer battery life and more human-like movement.

But hardware is only half the story. The AI models controlling these machines need vast quantities of high-quality physical data. Hyundai is generating it in factory environments, NVIDIA through simulation and Axonnic through muscle and force signals.

The next phase of the humanoid robotics race may therefore be decided not just by who builds the best robot, but by who builds the best data infrastructure around it.

News, insights, and features

Stay up to date with our latest thinking.