Physical Data Shortage Stalls Humanoid Robot Progress, Warns Telus Digital Executive

A Data Bottleneck Hits Humanoid Robotics

While rapid advancements in generative artificial intelligence over the past four years have accelerated the testing and rollout of humanoid robots, developers are now hitting an impasse caused by a severe shortage of training data. Speaking on the Targeting AI podcast, Sce Pike, vice president of AI growth and solutions at Canada-based Telus Digital, warned that the data pipeline that powered modern large language models cannot simply be copied over to physical machines.

While text-based models scaled rapidly by ingesting the public internet, physical robots cannot learn how to interact with the real world through text alone. As Pike, who leads Telus Digital's robotics and world model initiatives, pointed out, bridging the gap between digital reasoning and physical action requires entirely different data architectures that the industry is struggling to assemble at scale.

Compute Costs, Sensor Noise, and Safety in the Real World

A primary challenge lies in the mechanics of how machines understand physical reality. Pike cited an analogy popularized by former Meta AI chief scientist Yann LeCun, who noted that a four-year-old child possesses roughly five times the comprehension of an LLM because human children learn through sensory input and visual ocular nerves. By watching and interacting with their environment, toddlers organically master gravity, physics, and cause-and-effect—concepts that cannot be derived purely from written language.

To replicate this understanding in robots, developers rely on world models trained on real-world footage. However, capturing and utilizing this material introduces steep operational hurdles:

• Exorbitant Storage and Compute Costs: Collecting the sheer volume of high-resolution video needed to teach machines basic physics creates massive infrastructure and compute expenses.
• Sensor Fragmentation: Physical training data is gathered using a disjointed mix of lidar, infrared sensors, and various camera types, creating conflicting signals and substantial noise within datasets.
• High-Stakes Safety Risks: Unlike a chatbot where corrupted or poor data results in incorrect text, flawed physical data directly endangers human safety. Pike highlighted scenarios where erratic hardware behavior—such as a robot collapsing and flailing in a crowded commercial mall—presents unacceptable real-world liabilities.

The Push for World Models

The data crunch comes as the robotics industry shifts focus from pure mechanics to embodied intelligence. While conventional enterprise AI focuses on synthetic datasets and digital automation, roboticists are increasingly exploring foundation world models that can generalize actions with less raw data. Startups such as Nvidia-backed Skild AI, for instance, are attempting to bypass the data deficit by training robots to execute tasks from single video demonstrations. For now, however, overcoming the high cost of physical data collection and noisy multi-sensor integration remains the primary obstacle between laboratory prototypes and safe, autonomous deployment in public spaces.