Keep pulling the thread on Chelsea Finn.
Chelsea Fent believes the biggest risk for Physical Intelligence is technical failure due to the problem's difficulty, rather than competitive pressure.
Physical Intelligence's current robot policies are memoryless, operating only on the current image frame, and adding memory is a higher priority for the company than adding new sensor modalities.
Chelsea Fent believes the most important factor for achieving generalizability in robotics is increasing the diversity of robot training data, not just the quantity.
Physical Intelligence is building a large neural network model with the goal of controlling any robot to perform any task in any scenario.
Physical Intelligence has intentionally adopted an open strategy, which includes open-sourcing model weights, publishing technical papers, and sharing robot designs with hardware companies.
The RTX project demonstrated that a model trained on data pooled from different robot embodiments could often outperform models trained locally on a single robot type.
Chelsea Fent argues that while observational data of humans is valuable, it is not sufficient for training robots, which fundamentally require data from their own physical embodiment to learn motor control.
Research has shown that leveraging pre-trained vision-language models allows robots to perform tasks involving concepts, like identifying Taylor Swift, that were present in internet data but not in the robot's specific training data.
A major challenge in robotics is the lack of a large, pre-existing dataset analogous to Wikipedia for language models, which necessitates active data collection from real robots.
Physical Intelligence believes an open research culture is necessary to attract and retain the top-tier researchers and engineers required to solve general-purpose robotics.
Physical Intelligence prioritizes optimizing for data collection speed and diversity, leading them to use cheap robots with easy-to-develop teleoperation interfaces.
Unlike self-driving cars, robotics has many commercially viable use cases that do not require solving for an entire distribution of scenarios and have lower safety risks.