Galbot's Risk from Model Fragmentation

The core risk is that Galbot can ship early wins faster than it can unify them into a scalable robot brain. A stack split across GraspVLA, TrackVLA, and GroceryVLA works when tasks are tightly defined, like picking items, following movement, or running a store workflow, but every new use case then needs model handoffs, workflow glue, safety checks, and fresh training. As embodied AI shifts toward broader foundation models, that custom integration work becomes the real tax on expansion.

• Galbot is positioned as a commercial and industrial robot company, but its own model lineup is organized by task, not by one shared general policy. That means adding a warehouse, retail, or factory workflow is less like turning on a feature and more like stitching together a new operating recipe.
  1 sacra 2 sacra
• The market is moving the other way. NVIDIA is pushing GR00T as a general purpose humanoid foundation model and unified development stack, and Google DeepMind is extending Gemini into robotics, both aimed at letting one model reason across many tasks instead of training a separate policy for each one.
  5 nvidia 6 deepmind 7 nvidia
• Comparable companies are already being framed around data scale and generalization. Research on Figure, 1X, Foundation, and Generalist centers on collecting broad real world behavior data so robots improve across environments, which raises the bar for narrower systems that depend on per workflow tuning.
  3 sacra 4 sacra 8 sacra 9 sacra

The next phase of competition will reward robot stacks that can learn one skill and reuse it somewhere else. If generalist embodied models keep improving, the winners will be the companies that treat each deployment as fuel for a shared policy, not as another custom integration project.