Conference Presentation, Interview
Chelsea Finn: This is the State of the Art in Robotics
- Physical Intelligence aims to develop general-purpose robots capable of performing diverse real-world tasks, such as laundry or operating in unseen environments, within a two-year operational history.
- While user adoption trajectories for AI models may initially reach one million users within five days of launch similar to ChatGPT, physical AI distribution is expected to be slower due to hardware constraints.
- Future physical AI systems are predicted to require full autonomy and significantly fewer errors than current machine learning models, with reliability targets set at over 90% for tasks like making espresso and approaching 99% through automated scenario iteration.
- Data training challenges remain significant, as generating one million trajectories for a one-minute task requires approximately 700 robot days, though early human intervention to recover from dead ends and training general-purpose value functions on diverse experience videos aim to improve efficiency.
- The company plans to deploy the PI-07 model, which is expected to outperform fine-tuned specialist models in throughput and success rate immediately, marking a shift toward a "GPT-like era" characterized by compositional generalization and open-source policy development.
- Strategic directions include switching from per-site models to generalist policies for small teams, leveraging internet-scale real-world data mixes of teleoperated and autonomous experience, and integrating synthetic data generation by models attempting tasks.
- Capabilities are projected to reach the level of ChatGPT within the next few years, with future projects aiming for policies that exceed human teleoperation speeds and enable long-period autonomous reliability.
- Market dynamics may differ from the language model sector due to the high costs of embodied data and hardware, potentially concentrating advanced models in well-resourced labs, despite an optimistic outlook for an open-source robotics community and increased deployment of robots in the physical world.