Physical Intelligence

Weave Robotics reports deploying Physical Intelligence's pi0.6 on its Isaac robot at Sea Breeze Cleaners + Laundromat in San Francisco, autonomously folding shirts, trousers and shorts with some interventions. This named laundry deployment is distinct from the previously recorded warehouse partner workflow.

Physical Intelligence reports live partner deployments of its general-purpose robot foundation models, including robots operating in real customer warehouses and performing production tasks with human-in-the-loop recovery.

Physical Intelligence reports live partner deployments of its general-purpose robot foundation models, including robots operating in real customer warehouses and performing production tasks with human-in-the-loop recovery.

Physical Intelligence's FAST research documents action tokenization using a discrete cosine transform and byte-pair encoding to represent robot action sequences compactly. This specific representation-method contribution is independent of the existing broad pi-model evolution evidence.

Physical Intelligence has progressed from pi0 to pi0.7, adding open-world generalization, memory, online RL and steerable emergent capabilities across multiple robot embodiments and tasks.

Physical Intelligence has progressed from pi0 to pi0.7, adding open-world generalization, memory, online RL and steerable emergent capabilities across multiple robot embodiments and tasks.

Physical Intelligence announced its Pi API beta on October 13, 2026, describing company-managed fine-tuning compute and hosted inference. Its account of early access names Almond, whose robot ran a model after four days of API iteration, and reports cloud-hosted use across initial applications. This supports a delivered model-service operating capability, distinct from the existing generic investor and partner description.

Physical Intelligence is building a reusable physical-intelligence layer, works with deployment partners, open-sources model weights and code, and is backed by investors including OpenAI, Sequoia, CapitalG, Thrive and others.

Physical Intelligence is building a reusable physical-intelligence layer, works with deployment partners, open-sources model weights and code, and is backed by investors including OpenAI, Sequoia, CapitalG, Thrive and others.

Physical Intelligence's real-time chunking research describes generating a next action chunk asynchronously while the current chunk executes, preserving committed actions through inpainting. The latency-handling technique offers transferable robotics/control-system engineering beyond existing broad role descriptions.

Work spans robot foundation models, reinforcement learning, multimodal learning, software engineering, hardware, robotics and general-purpose embodied AI across multiple robot platforms.

Work spans robot foundation models, reinforcement learning, multimodal learning, software engineering, hardware, robotics and general-purpose embodied AI across multiple robot platforms.