Shortage of training video data is holding back humanoid robot development, Telus Digital warns
According to Sce Pike of Telus Digital, the development of humanoid robots is being held back by a shortage of training video data, its high storage and processing costs, and noise from diverse sensors - which also affects the physical safety of robots.
Sce Pike, Vice President of AI Growth and Solutions at Telus Digital, described the shortage of training video data as one of the main brakes on humanoid robot development in the Targeting AI podcast (AI Business). Pike leads projects at Telus Digital focused on robotics and world models. According to her, language models were able to draw on virtually the entire internet, whereas physical AI has no such data source available to it.
Pike referenced the well-known comparison made by former Meta AI chief scientist Yann LeCun, according to which a four-year-old child has roughly five times greater understanding of the world than today's language model. According to LeCun, a child learns about physics, gravity, and cause and effect through visual and sensory perception, which cannot be replaced by merely reading text.
According to Pike, one of the main obstacles to training advanced robots and the world models that control them in the real world is the enormous volume of video recordings and the associated prohibitive costs of computing and storing the data. Another problem is the inconsistency of data from different sensors - cameras, lidars, and infrared sensors - which, in her words, creates significant noise in the data.
Pike also deals with the physical safety of robots and pointed out that poor training data has far more serious consequences in the physical world than with chatbots. While a chatbot, in the worst case, provides incorrect information, a robot with insufficiently trained behavior can, for example, fall over and move uncontrollably in an environment with people, such as in a crowded shopping center.
Why it matters
The piece describes why physical AI cannot be trained the same way as language models - robots lack the equivalent of the internet's text corpus in the form of video and sensory data, which, according to Pike, slows development and increases the risk of accidents when robots are deployed among people. For companies developing robots and world models, this is a signal that investment in collecting, storing, and unifying sensory data is a key factor for both safety and development speed.
Relevant practical impact
What this means
For a business
Companies developing humanoid robots and world models are running into a shortage of training video data and the high costs of storing and processing it, which may extend development timelines and increase the costs of achieving safe deployment of robots in real-world environments.
Development More business impacts →Check the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.