π-0.5, a model doing long-horizon tasks in real, unseen homes with unseen objects! Physical Intelligence one of my favorites below at 10x sharing a few of my favorite results in thread, with many more detailed ablations in the paper:

with ~100 ood environments in the training data, we match or exceed in-distribution performance (where in-distribution had hundreds of hours) further, we see that without pretraining, the ood performance drops substantially

a surprising result to me was how much co-training with high-level outputs helped performance. the model implicitly learns to break down tasks, though an explicit high-level policy is still better HL demos to learn the performance of the LL was also critical

for language following: multi-environment, cross-embodiment, and web data were all critical to ood performance

zooming in. as we scale environments, language following improves too, and particularly recognizes objects lags being able to manipulate them

@physical_int Imagine u make a complete system that gets all the data we humans get during our every day life. Especially the physical part Like we use sensors etc And then we do that at scale with millions of participants and we use that data to train a super intelligent self aware system