The Weakness of Robots Is Not the "Hardware"—What Was Missing Was the "Brain"
The future where robots "do the dishes at home, assemble boxes, and even brew coffee" cannot be reached through hardware evolution alone.Karol Hausman and Tobi Springenberg of Physical Intelligence (π)assert that the true bottleneck in robotics is "intelligence." In fact, they are building a robot version of a "foundation model" from the position that "the bottleneck for robots was intelligence, not hardware," and are advancing from imitation learning to reinforcement learning through field experience.
1. Why "decomposed" robotics has reached its limit
Conventionally, the standard approach has been to decompose robots into perception, planning, and control, designing and training each module separately. However, they view this decomposition as the very core of "why it doesn't work." When humans pick up a cup, we don't feel like we are sequentially processing "recognize, plan trajectory, control..."—our actions are much more integrated. Therefore, it is more natural to learn everything from pixels (vision) and instructions (language) to actions in one go.
2. What is a robot "foundation model"?
2-1. π0: A "VLA" that connects images, language, and actions
What π proposes is a Vision-Language-Action (VLA) model.In short, it is a design that attaches an action expert to move the robot onto a backbone (VLM) that understands images and language, outputting even low-level motor commands. π0 (pi-zero) is their first generalization policy, clearly establishing a worldview where "if the user speaks, the robot learns the action."
2-2. Why a model company, not a "siloed robot company"?
Hausman and his colleagues' argument is simple. Robots have been able to do many things for a long time "if operated by a human." In other words, the ceiling (hardware potential) has been rising, but the floor (autonomous intelligence) is low. Therefore, "clearing the intelligence bottleneck first" becomes the foundation that can expand vertically into homes, factories, medicine, and more.
3. The hardest part is "operating in a new home"
They divide the challenges into Capability, Generalization (adaptation to new environments), and Performance (practical stability). Generalization, in particular, is prone to failure in "new homes" where room layouts, lighting, and object placement are different. π0.5 has stepped into this generalization, testing long tasks like cleaning up kitchens and bedrooms in homes not in the training data, showing a certain level of flexibility (though still incomplete).
What is important here is their pragmatic approach. The only known prescription for generalization is ultimately "diverse data." Therefore, they increase the "variance" of homes, objects, and tasks to raise the probability of extrapolating to similar situations.
4. "π*0.6" and reinforcement learning (RL) that broke the wall of imitation learning
4-1. Failures chain—why imitation learning is weak in the field
Because robots touch the real world, small deviations change the next state, and errors snowball. With only imitation learning, this "chain" becomes a fatal wound. π addresses this head-on, stating that practical application requires a "mechanism to improve from field experience."
4-2. RECAP: Using experience + human corrections to become "2x faster and 2x more stable"
π's π*0.6 uses RECAP (RL with Experience & Corrections via Advantage-conditioned Policies), which improves through the flow of:
creating an initial policy via demos (human teleoperation)
collecting experience by running it in the field
having humans provide "success/failure" feedback or "fix it like this" instructions
. As a result, they report improvements such as throughput more than doubling and failure rates being halved in some tasks.
Symbolic of this is the "duration" of the demos. For example, they showed a picture of "operations" rather than "one-off tricks," such as making espresso from morning to night, folding various laundry in a new home for several hours, or running a cycle of assembling chocolate boxes and labeling them in a factory.
5. What This Means for the "Next Competition"
Summarizing their points from an editor's perspective, the competition is happening on three layers.
Foundation Modeling: Learning pixels to actions as a single unit (a shift in design philosophy)
Generalization: Conquering "unseen homes" using diverse real-world data
Growth Through Operation: Surpassing the limits of imitation learning with experience-based RL
And commercially, as they themselves say, "we haven't decided how to sell it yet," the final form could be either model provision or vertical integration. However, the direction is clear: "Once it reaches a level of performance that can be deployed, data will return from the world, and the model will become even smarter with that returned data." Whether this self-amplifying loop can function will likely determine the "next OS" of the robotics industry.