Why Robots Still Struggle With Simple Tasks (And What Might Finally Change That) | Karol Hausman, Co-Founder & CEO of Physical Intelligence

Physical Intelligence bets generalist robot models will beat specialists

Robots have failed for decades because every task had to be hand-programmed, and the physical world's complexity cannot be written down as rules. Physical Intelligence bets the recipe behind language models, one generalist model trained on diverse real-world data and scaled, will work for manipulation, and that reinforcement learning can now codify task success instead of copying actions. Chasing an early application is what turns robotics companies into pick-and-place vendors, so the company optimises for rate of learning.

Takeaways

• Generalist robot models will beat specialists. Hausman compares robotics to language, where the best translation and coding systems came from training one generalist model rather than many specialists. The models absorb diverse data, including videos of people, teleoperation, handheld devices and simulation.
• Manipulation needs real data, not simulation. Simulation works for locomotion and backflips, where you only model your own body. Picking up a cup requires modelling how the world reacts, and getting friction parameters right for every object and task does not scale.
• Reinforcement learning returns to codify success. Imitation learning optimises for copying actions rather than task success. Reinforcement learning can encode success directly, and foundation models like Pi Zero solve the exploration problem by not starting from scratch, targeting 99.99% reliability rather than 70%.
• Commercialising early makes robots pick-and-place vendors. Hausman calls chasing the biggest-TAM application the classic robotics failure. Revenue, customers, valuation and employee incentives push a company to cut corners on technology until it becomes a special-purpose application company. Physical Intelligence optimises for rate of learning instead.
• First deployments came 18 months in. The company expected five years. Pi Zero folded laundry within seven months. Pi 0.5 worked in homes it had never seen. Pi 0.6 made espresso for 13 hours straight and cleaned up after itself.
• Value functions spot failure before humans do. The physical world is unforgiving. Hausman cites AlphaGo against Lee Sedol, where commentators thought the game was close while the value function showed the AI had already won with certainty.
• Wrist cameras and intelligence beat precision hardware. Physical Intelligence put wrist cameras on its robots instead of touch sensing, and the models compensate, perhaps by seeing deformation on the gripper that indicates pressure. Its robots are imprecise next to industrial machines, but intelligence covers the gap.

From a Polish town to Physical Intelligence