Physical AI shifts toward general-purpose embodied foundation models

GroundingPI reports spatial pretraining and gains across 34 grounding benchmarks, while Natura-1 claims a 50B-parameter robot-hand model with zero-shot manipulation and 100-ms control. AlphaNIST adds a simulation-based precision-assembly signal but emphasizes perception, latency, contact mismatch, rollout cost, and sim-to-real barriers; company claims and benchmark results still need independent real-world validation, with GR00T N1.7’s 56% household-task completion underscoring the reliability gap.