Is physical AI ready for real work?

#	Model	Benchmark Single-arm pick-and-place DROID · v1.0
		Success @ 30 s	Time per item
-	Robot teleoperated by Human Teleoperation Baseline	99.0% ± 0.4	10.4 s ± 0.2
1–2	NVIDIA GR00T N1.6 Generalist VLA Finetuning	46.1% ± 3.9	15.7 s ± 0.7
1–2	Physical Intelligence Open π₀.₅ Generalist VLA Finetuning	44.4% ± 4.3	15.7 s ± 0.8
3	Action Chunking Transformer Specialised	42.5% ± 5.6	13.8 s ± 0.9
4	Hugging Face SmolVLA Generalist VLA Finetuning	16.3% ± 4.5	16.3 s ± 1.6

Success @ 30 s is the share of items placed within 30 s of the previous item. Time per item is the mean time of those items. The rank orders the models by Success @ 30 s. A rank such as 1–2 means that chance can explain the difference between the time-to-success curves of the models in those places. ± is half the 95% range. Select a row to open the model's page.