A simple example of physical prompt steerability: same environment, different prompts, different behaviors. Thanks Jagdeep Bhatia for the suggestion. Read more about GEN-1.5 in our blog post in the comments below.

Other tasks our models can do: Read more about GEN-1.5, our latest foundation model for the physical world:

@JagdeepBhatia8 Very cool!

@JagdeepBhatia8 What if one re-stack those 3 blocks a few times, e.g. (red, blue, green) -> (green, red, blue) -> (blue, red, green) -> ... see if robot re-stacks them in same order at every step. Bonus points if number of re-stacks is more than number of layers of model.

@JagdeepBhatia8 这是个诈骗公司

@JagdeepBhatia8 Testing the same scene with different physical prompts is a much cleaner way to separate prompt following from ordinary generalization. How much behavioral variation can GEN 1.5 sustain before the prompt falls outside its training distribution?

@JagdeepBhatia8 What happens if you interpolate between two physical prompts in the same scene? Does behavior change gradually or jump between modes?