Why humanoid robot demos still fail the generalization test

Appen's Janine Sinanan-Singh argues impressive humanoid robot demos mostly show narrow trained skills, not true generalization, and still require human-in-the-loop data quality control.

• Demos don't prove generalization: Appen's Janine Sinanan-Singh says humanoid demos (e.g., folding laundry) can't show if skills transfer to out-of-distribution tasks, questioning claimed "zero-shot" wins.
• Foundation models speed robotics bootstrapping: Plugging general multimodal models like Gemini or Opus into robots can complete basic tasks without fine-tuning, though with ~7-minute latency versus purpose-trained models.

Themes, sentiment and reasons are AI-written paraphrases, checked so they never quote an episode. They may contain errors. Not investment advice. Episodes remain the property of their creators.