Sign up to save tools and stay up to date with the latest in AI
bg
bg
1

π0: Our First Generalist Policy

Nov 10, 2024 - physicalintelligence.company
Physical Intelligence has developed a general-purpose robot foundation model called π0 (pi-zero) that aims to enhance the physical intelligence of AI systems. The model, trained on diverse data, can follow various text instructions and control different robots, learning to output low-level motor commands through a novel architecture. The model can perform a wide range of tasks, from folding clothes to assembling boxes, and can be fine-tuned for specific application scenarios.

The company compared π0 to other robot foundation models and found it to be the best-performing model across all tasks. Despite the promising results, the company acknowledges that generalist robot policies are still in their infancy and there are challenges to overcome, including long-horizon reasoning and planning, autonomous self-improvement, robustness, and safety. Physical Intelligence is seeking collaborations with the robotics community to further refine the model and is also open to hiring new talent.

Key takeaways:

  • The team at Physical Intelligence has developed a general-purpose robot foundation model called π0 (pi-zero), which is a step towards developing artificial physical intelligence. This model is trained on broad and diverse data and can follow various text instructions, control a variety of different robots, and can be fine-tuned to specialize in challenging application scenarios.
  • π0 uses Internet-scale vision-language pre-pretraining, open-source robot manipulation datasets, and the team's own datasets consisting of dexterous tasks from 8 distinct robots. The model can perform a wide variety of tasks, including folding laundry, bussing tables, and assembling boxes.
  • The model inherits semantic knowledge and visual understanding from Internet-scale pretraining by starting from a pre-trained vision-language model (VLM). To enable high-frequency dexterous control, the team developed a novel method to augment pre-trained VLMs with continuous action outputs via flow matching, a variant of diffusion models.
  • The team at Physical Intelligence is looking to collaborate with companies scaling up data collection with robots deployed for real-world applications, who are looking to collaborate on autonomy. They are also hiring and encourage interested individuals to get in touch.
View Full Article

Comments (0)

Be the first to comment!