Combining human-in-the-loop intervention to avoid dead-end trajectories with a general value function trained on diverse robot experience enables 2x throughput gains and 90%+ success rates…

A scalable RL recipe combining early human intervention to avoid dead-end trajectories and a generalizable value function trained on diverse robot experience doubles task throughput and ach…

Combining short-term video memory (10s) with long-term compressed text summaries (10-15min) enables robots to execute non-repetitive, multi-step tasks like kitchen cleaning autonomously, so…

A memory architecture combining short-term video memory (10 seconds) with long-term textual summaries (10-15 minutes) enables robots to execute multi-step, non-repetitive tasks like kitchen…

Physical Intelligence's PI07 model matches or exceeds fine-tuned specialist models across diverse tasks (espresso making, box building, folding) without task-specific training, marking a 'G…

Physical Intelligence's π0.7 model achieves out-of-the-box performance matching or exceeding task-specific RL-finetuned specialists across diverse manipulation tasks, marking a transition f…

π0.7 demonstrates Dolly-like compositional generalization: operating unseen appliances (air fryers) and transferring folding skills to a completely different bimanual robot platform without…

Autonomous robot deployment data — not human video — is the critical training fuel; PI0/PI05 open-source releases already enable small teams to fine-tune generalist policies, lowering barri…