After the walking demo, the next question is: what will the robot learn from?

Humanoid competition in 2026 is becoming harder to explain through hardware demos alone. A robot that must handle unfamiliar objects, adapt a task sequence and recover from failure needs repeated experience of the physical world. Capability therefore depends not only on the model, but also on how diverse experience is created and how quickly that experience can cycle back into training and evaluation.

That changes the labor map. Robotics companies need more than AI researchers. They need people who connect real-world data collection, simulation environments, quality review, distributed training infrastructure, policy evaluation and deployment to hardware. Recent facilities and job postings show this middle layer becoming visible as an organization of its own.

ACE’s “roughly 100,000 hours” is not an official industry statistic — but it is a signal of the perceived data gap

In a Reuters interview published on August 21, 2026, ACE Robotics chairman Wang Xiaogang estimated that current robot-training datasets amount to roughly 100,000 hours. ACE also said it plans to collect tens of millions of hours from production environments within two years.

The figure should not be treated as a verified total for the global robotics industry. Public and proprietary datasets cover different scopes, and companies do not necessarily count “training hours” in the same way. What matters for talent analysis is that a robotics company identifies data scarcity as a core constraint and plans to expand collection by orders of magnitude. That changes where capital and people are likely to be allocated.

Apptronik is building a factory for robot experience, not just a factory for robots

Apptronik’s Austin Robot Park, announced in June 2026, is a nearly 90,000-square-foot data-collection and training facility. Apollo 2 fleets perform logistics, manufacturing, retail and other tasks, while similar collection workflows extend into customer and partner sites.

The collection model is hybrid. Apptronik describes a combination of teleoperation, autonomous execution and high-fidelity physics simulation. Humans demonstrate, robots execute, success and failure are recorded, simulation expands the experience, and the data returns to model training. In that sense, Robot Park functions as a factory for embodied experience.

Figure’s job board shows robot data already splitting into distinct professions

Figure is currently hiring an AI Training Infrastructure Engineer for humanoid whole-body control. The role owns simulation, data pipelines, orchestration and cluster utilization, and builds the backbone that moves policies from training to validation and deployment on hardware. The job combines Python and PyTorch with physics simulation, controls, robotics systems and distributed infrastructure.

At the other end of the pipeline, Figure’s Data Collection Quality Support role reviews collection sessions, identifies inconsistent technique, incomplete sequences and edge cases, and produces feedback that directly shapes robot training. One role looks like high-performance ML infrastructure; the other looks closer to data QA and operations. Both sit inside the same system for turning robot activity into useful learning data.

When real-world data is expensive, simulation and synthetic data become a second production line

Collecting every failure and rare condition by operating physical robots is slow and expensive. NVIDIA’s Physical AI Data Factory Blueprint frames the problem as a pipeline of data curation, synthetic generation, reinforcement learning, evaluation and orchestration. Limited real data can be expanded into rarer and longer-tail scenarios, then filtered and validated before training.

NVIDIA says GR00T 1.7 was pretrained on about 32,000 hours of real demonstrations and egocentric data plus roughly 8,000 hours of simulated rollouts and demonstrations. The more important point is the mix. Robot-learning teams increasingly have to combine real sites, teleoperation, simulation and synthetic generation — and understand how each source changes performance on hardware.

For companies, this becomes a data-production system design problem, not simply an “AI hiring” problem

For a company trying to commercialize Physical AI, data becomes operating infrastructure rather than a by-product of research. Teams must decide which tasks to collect, who demonstrates them, how sensor and robot-state data are stored, how failures are classified, where simulation connects to the physical system, and which evaluation metrics allow a policy to move into deployment.

Hiring requirements follow that architecture. Companies may need people who bridge MLOps and distributed systems, simulation, robotics controls, data QA and field deployment. More collection alone is not enough: if quality criteria and evaluation are weak, a larger dataset can still fail to accelerate learning.

For workers, new adjacency paths are opening from careers that did not used to be called robotics

This is not limited to robotics majors. Engineers who have operated large-scale ML training pipelines may be adjacent to robot training infrastructure. Game, CAE and digital-twin specialists may map into physics simulation. Manufacturing quality or data-review experience may connect to collection quality, while equipment commissioning can overlap with data-collection deployment and operations.

But adjacency is not equivalence. Cloud experience alone does not make someone a robotics engineer. Figure’s training-infrastructure role still asks for familiarity with dynamics, controls and robotics systems. The career question is therefore not “Can I rename my job?” but “Which part of my existing experience can be evidenced inside the robot data creation, training, evaluation or deployment loop?”

Data can be a major bottleneck without replacing hardware, safety and economics as constraints

The evidence does not support the claim that data is the only bottleneck in humanoid robotics. Actuator reliability, batteries, safety, manufacturing cost, customer ROI, task speed and maintenance all remain critical to commercialization. Recent market discussion is also shifting from spectacular demos toward productivity, autonomy and economics in real work.

The narrower conclusion is stronger: across leading companies in 2026, the ability to create data at scale, control its quality and repeatedly move it between models and real robots is becoming an independent competitive axis. As that axis expands, the jobs around it expand too.

BANSEOG VIEW

Physical AI talent competition is expanding from people who build robots to people who build robot experience

Counting “robotics engineers” alone misses part of the emerging labor market. A learning robot depends on people who create demonstrations, scale simulation, judge data quality, operate large training jobs and push validated policies onto hardware. Together they form something closer to a learning factory than a conventional AI team.

Banseog Search views this less as a list of new titles and more as a map of career adjacency. MLOps, simulation, manufacturing QA, equipment deployment and controls already exist in other industries. The next talent question is where those experiences can credibly connect to the Physical AI data pipeline.

SOURCES

Primary sources and references

ACE’s roughly 100,000-hour estimate and tens-of-millions target are figures reported by Reuters from ACE Robotics and are not an official census of global robot data. Apptronik and NVIDIA are company primary sources, while Figure postings describe one company’s current roles. We therefore do not generalize these sources into an industry-wide hiring count or salary premium. The conclusion that data production, training, evaluation and deployment are becoming a separate competitive layer is Banseog HR Intelligence’s synthesis across these signals.