Thinking Machines: Embodied Intelligence (10-minute read)

The Pitfalls of Data Scarcity and Generalization

Despite advancements in VLA (Vision-Language-Action) models, the robotics field faces limited growth due to the lack of rich training data corpora, unlike LLMs. A structural bottleneck exists where actual robots and teleoperators must generate data at physical speeds. Analysis of the Libero benchmark reveals that models have not learned manipulation but rather memorized specific trajectories, resulting in a complete failure pattern where success rates plummet from 95% to below 30%, or even 0%, with minor environmental perturbations.

The Reality of Industrial Robots and Data Assetization

The 4.66 million industrial robots currently in operation are merely traditional automation programmed for specific jigs and parts, lacking generalization capabilities. As seen in the deployment of BMW's Figure 02, humanoids remain in a high-cost experimental phase, operating for less than four hours on average per day. Conversely, this installed base represents the largest corpus of physical interaction data, left uncollected in a learnable format. Instrumenting this data to convert it into training data could become a core asset that is impossible to replicate with VC funding.

Reliability Standards and Future Outlook

Systems interacting with the physical world require 99.9% or higher accuracy and continuous evidence of safety, unlike language models. In environments where probabilistic behavior is unacceptable, errors lead to catastrophic losses. The author diagnoses robotics as a data problem rather than a model problem, emphasizing the necessity of accumulating data in narrow, well-defined domains. True Embodied Intelligence will only emerge after such data-driven reliability is secured.