Physical Intelligence, a San Francisco-based startup, is at the forefront of developing general-purpose robotic intelligence, often likened to "ChatGPT but for robots." Their headquarters, subtly marked by a variant pi symbol, bustles with activity, featuring a giant concrete box layout with long blonde-wood tables. These tables serve dual purposes, some for meals, adorned with snacks like Girl Scout cookies and Vegemite, while many others are laden with monitors, spare robotics parts, wires, and fully assembled robotic arms. These arms are engaged in mastering mundane tasks, such as attempting to fold pants, turn shirts inside out, and efficiently peel zucchini, showcasing the company's testing phase.

Sergey Levine, an associate professor at UC Berkeley and a co-founder of Physical Intelligence, explains that this observed activity is part of a continuous loop. Data collected from these robot stations, as well as from various other locations like warehouses and homes, is used to train general-purpose robotic foundation models. Once a new model is trained, it returns to these stations for evaluation, with each robotic task serving as an experiment. For instance, the zucchini-peeler might be testing the model's ability to generalize peeling motions across different vegetables it has never encountered. The company further extends its testing by operating a test kitchen, exposing robots to diverse environments and challenges using off-the-shelf hardware, even utilizing a sophisticated espresso machine as a data source for robots to learn, rather than a staff perk.

The hardware itself is intentionally unglamorous; the robotic arms cost around $3,500, with an "enormous markup" from the vendor, potentially dropping below $1,000 if manufactured in-house. This deliberate choice underscores the company's core philosophy: superior intelligence can compensate for less advanced hardware. Lachy Groom, another co-founder and a prominent Silicon Valley figure who made early investments in companies like Figma and Notion, was drawn to Physical Intelligence after closely following the academic work of Levine and Chelsea Finn, a former Berkeley PhD student now leading her own lab at Stanford focused on robotic learning. Groom's pursuit led him to Karol Hausman, a Google DeepMind researcher, confirming his conviction that this was the right company to commit to after a five-year search for his next venture post-Stripe.

Despite raising over $1 billion and being valued at $5.6 billion by backers including Khosla Ventures, Sequoia Capital, and Thrive Capital, Groom maintains an unusual stance with investors: he provides no timeline for commercialization. He notes that most of the company's spending goes towards compute, and they are always open to raising more capital as "there's no limit to how much money we can really put to work." Quan Vuong, another co-founder from Google DeepMind, elucidates their strategy, which centers on cross-embodiment learning and diverse data sources. This approach aims to reduce the marginal cost of onboarding autonomy to any new robot platform, enabling the transfer of existing knowledge without starting data collection from scratch. Physical Intelligence is already collaborating with a select number of companies in various sectors, such as logistics, grocery, and a chocolate maker, to test the real-world applicability of their automation systems, with some applications already proving effective.

The race to build general-purpose robotic intelligence is intensifying, with Physical Intelligence facing competition from entities like Pittsburgh-based Skild AI, founded in 2023, which recently raised $1.4 billion at a $14 billion valuation. Skild AI has taken a notably different path, commercially deploying its "omni-bodied" Skild Brain and generating substantial revenue in security, warehouses, and manufacturing. This highlights a sharp philosophical divide: while Physical Intelligence prioritizes pure research, betting that resisting near-term commercialization will yield superior general intelligence, Skild AI believes commercial deployment creates a data flywheel that continuously improves its models with real-world use cases. Skild AI has publicly criticized competitors, arguing that many "robotics foundation models" lack "true physical common sense" by relying too heavily on internet-scale pretraining rather than physics-based simulation and real robotics data. The ultimate resolution of which approach is "more right" is expected to take years.

Physical Intelligence operates with what Groom describes as "unusual clarity," driven by researchers' needs to collect data, develop new hardware, or address specific challenges without external pressures. The company rapidly surpassed its initial 5-to-10-year roadmap within 18 months. With approximately 80 employees, the company plans for controlled growth, acknowledging that hardware presents the most significant challenges due to breakage, slow delivery, and safety considerations. Despite external skepticism regarding the demand for household robots, safety concerns, and the viability of their long-term vision, Groom remains confident, backed by decades of experience from his team, who believe the timing for this technological leap is finally right. Silicon Valley's historical support for such ambitious ventures, often without immediate commercialization timelines, reflects a belief that pioneering teams will ultimately find solutions, justifying significant investments even amidst uncertainty.