Startup Physical Intelligence unveils π0.7, an AI model for robots that performs tasks without specific training, showcasing potential for universal robotics.
Physical Intelligence, a startup established by ex-Google engineers, has launched its latest AI model, π0.7. This model is touted as a significant advancement in the ability of artificial intelligence to generalize skills and execute tasks without needing specific training.
"Our newest model, π0.7, demonstrates remarkable emergent capabilities, such as controlling a new robot to fold shirts despite having no prior data for this task," noted the team at Physical Intelligence.
This new AI system, categorized as "Vision-Language-Action" (VLA), is specifically designed for robot operation. Unlike its predecessors, π0.7 showcases compositional generalization, allowing it to blend learned skills to manage new tasks effectively.
Untrained Tasks and Robot Adaptability
During testing, π0.7 was able to manipulate a novel robot type to fold t-shirts, a task it had not been explicitly trained for. The performance was comparable to that of human operators with extensive experience in remote control.
The AI also displayed an ability to navigate unfamiliar kitchen devices, completing tasks like preparing sweet potatoes in an air fryer without prior exposure to similar scenarios.
According to the developers, this capability stems from the model’s integration of various skills, akin to how language models utilize knowledge from diverse fields.
Versatile Control Mechanisms
One of the standout features of π0.7 is its command-processing ability, which includes not only direct instructions but also contextual guidance. It can accept:
- text-based directives;
- execution quality and speed parameters;
- visual prompts that illustrate intended outcomes.
This flexibility enables the AI to adapt its actions during operation without needing additional training sessions.
"π0.7 can respond to a variety of prompts that convey not just 'what' to do, but also 'how' to do it, integrating prompts from multiple modalities," the team highlighted.
Towards Universal Robotics
Physical Intelligence pointed out that earlier models required retraining for new tasks, similar to the limitations seen in early language models. π0.7 distinguishes itself by functioning "out of the box" and adapting to new situations through language commands.
While the model shows impressive capabilities, it still requires detailed step-by-step instructions for more complex tasks; however, performance improves significantly with structured guidance.
The company envisions a future where such models could lead to more autonomous robots that operate independently without needing manual configuration for each new task. π0.7 marks a pivotal step toward creating adaptable robots capable of responding to varied conditions.
