I have high respect of Tuomas and his work around SAC for RL in robotics.

But this is slightly unconvincing, most because of the author

>They spend thousands of computer hours practicing movements inside simulated worlds and inventing their own solutions.

This is exactly what almost every other picking startups have been doing for the last couple of years.

I can think of at least a dozens, some even making their custom gripper hardware. It still relies on sim2real transfer and then there's a bottleneck of things such as representing deformable objects. And that's still just scratching the surface of it.

I can definitely see that they have the right team. But the claim made by this author is far removed from the actual demo he describes. I've seen same demo for years, last one was in CoRL by Google (Gemini) and even then you could see clever robotics guy (some Boston dynamics engineers) that came by and gave it a clever task it failed on.