Please keep in mind that "uncut footage of ..." is the same as those instagram videos where people throw a basketball backwards 20 meters while drinking beer, straight into the goal.

It doesn't prove this is average performance. In fact, if it was, that's not what they'd be showing.

From my skim of their paper, they have gotten robodojo 14% versus previous sota 9%. So its an improvement, but doesnt seem like a reliable model yet, which is fine. I also notice the model becomes worse with more data on the printer refilling task, so thats either a bad run or its a data point against the robustness. The transfer learning results are meaningful. 75% on new complex tasks from <10 hours of demo is really good. I believe this is a 2x improvement over the SOTA. 10 trials per task is standard for these papers but is pretty weak statistically. Memory tasks its weaker at but they have acknowledged it, and I cant see why this arch would have problems w that in the future.