Physical IntelligencePhysical Intelligence

About

Physical Intelligence (pi) was founded in San Francisco in 2024 by a cluster of the most-cited researchers in robot learning — former Google DeepMind Senior Research Scientist Karol Hausman, UC Berkeley professor Sergey Levine, Stanford professor Chelsea Finn and others — to 'build general AI for the physical world': it builds no robot bodies and focuses on general robot foundation models that can control any robot to perform any task, often called the 'OpenAI of robotics'. Its core products are the pi-series vision-language-action (VLA) models: pi0, the first general policy (a VLA Flow Model), released in October 2024; open-sourcing of the pi0 and pi0-FAST weights and code (openpi) in February 2025; pi0.5 with open-world generalization in April 2025 (able to operate mobile manipulation robots to tidy kitchens/bedrooms it has never seen); pi*0.6 in November 2025, learning from experience via reinforcement learning; and pi0.7 in April 2026, a steerable model with a significant leap in generalization. Funding: a ~$70M seed round in March 2024 ($400M valuation); a $400M Series A in November 2024 ($2.4B post-money, with Jeff Bezos, OpenAI, Thrive Capital, etc.); a $600M Series B in November 2025 (led by CapitalG, $5.6B post-money), totaling over $1.07B; and in March 2026, per Bloomberg, it was in talks for ~$1B at a valuation possibly above $11B (unverified).

Technical Approach

Technical route: VLA models at the core — pi0 generates continuous actions via flow matching, with pre-training that fuses internet-scale vision-language data and large-scale multi-robot real manipulation data; later versions evolve toward 'stronger generalization + learning from experience' — pi0.5 achieves open-world generalization, pi*0.6 introduces online reinforcement learning to improve real-task success rate and throughput from real interaction, and pi0.7 emphasizes steerability and emergent capabilities. Supporting component research includes the FAST action tokenizer (5x faster training), HiRobot hierarchical language control, MEM short/long-term memory, and RL Token for efficient online RL. It commits to open source (the openpi repo supports PyTorch and offers DROID dataset training recipes); CEO Karol Hausman advocates reusing the LLM scaling path — one model generalizing across bodies and tasks, with the data flywheel as the core moat.

Commercialization（4）

• ◆Cumulative funding over $1.07B from investors including Jeff Bezos, OpenAI, Thrive Capital, Lux Capital, Bond, Khosla Ventures, Sequoia, CapitalG (Alphabet), Redpoint, Index Ventures and T. Rowe Price
• ◆In February 2026 published 'The Physical Intelligence Layer', saying partners are already solving real-scenario problems with its general models and that this will trigger a 'Cambrian explosion' of robot applications
• ◆The openpi open-source ecosystem: pi0/pi0-FAST weights and training code open, PyTorch support, and full DROID dataset training recipes
• ◆Team of about 80 (per March 2026 media reports); no explicit commercialization timeline yet, prioritizing model capability and data scale

Position & Strengths（5）

• ◆The founding team is one of the most-cited researcher clusters in robot learning (Levine, Finn, Hausman and Ichter all have landmark works)
• ◆The pi-series VLA models iterated 5 versions in 18 months (pi0 → pi0.5 → pi*0.6 → pi0.7), with industry-leading architecture evolution and data flywheel speed
• ◆A single model trained across many robot bodies; models only, no bodies — avoiding head-on competition with robot OEMs
• ◆openpi open source builds a developer and researcher ecosystem while keeping the latest versions ahead
• ◆Top capital endorsement: Bezos, OpenAI, CapitalG, Thrive and others; valuation grew from $400M to $5.6B in two years (rumored $11B in a new round)