Physical Intelligence

Multi-scale Embodied Memory keeps short-horizon video history and long-horizon language summaries that the model chooses to store, enabling tasks up to 15 minutes long and in-context recovery from earlier mistakes.

RL tokens summarize the VLA’s internal state into a compact interface, letting a tiny on-robot actor-critic learn online and speed up the most precise phases of tasks by up to 3x from about 15 minutes of real data.

Hi Robot is a two-level system: a high-level VLM breaks complex prompts and live corrections into single steps "whispered" to the π0 low-level policy, trained partly on synthetically labelled interactions.

Shows that simply co-fine-tuning a scaled VLA on egocentric human video transfers to robot behavior, and that this ability emerges as robot pretraining data grows, giving up to 2x improvement where robot data is scarce.

Knowledge insulation stops gradients from newly added continuous action experts from overwriting the VLM backbone, so the model trains as fast as tokenized VLAs while keeping web-scale semantic knowledge.

Real-time chunking lets diffusion- and flow-based VLAs generate the next action chunk while the previous one is still executing, with no training-time changes and without the discontinuities that break naive smoothing.

A roughly 3B-parameter vision-language-action model that adds a flow-matching action expert to a pretrained VLM and is fine-tuned to dexterous multi-robot tasks such as laundry folding and table bussing.

Compresses continuous robot action chunks with a frequency-domain tokenizer so vision-language-action models can be trained autoregressively at high speed without losing dexterity.