[draft]

Some aspects of a useful robotics model:

  • Accurate: a fast and consistent base model, with in-weights training to get it to full accuracy.1
  • Responsive: 10 ms of input lag makes a game unplayable and makes an AI unable to have feedback in the real world.
  • Affordable: intelligence will live on every device; as the cost of training and running the models goes to zero1, the largest cost will be hardware spending; we aim to make the most affordable device per unit of capability.
  • Convenient: the main cost of hardware compared with software is that it takes up space; we need to replace or be part of existing household spaces. We do this by replacing the traditional hamper/laundry basket.
  • Fun: the difference between Microduck and Reachy Mini is immense; they are similar in capability, yet one is more relatable, more human, and customizable with open-core roots.

Footnotes

  1. 2 parts - model = closed weights plus an open-weight fine-tune; you may literally own the weights, or it will be fine-tuned via API; or models will be local; fine-tuning is required when intelligence can go on any device. ↩ ↩2