Physical Intelligence, an artificial intelligence company based in San Francisco, has published new scientific research detailing a revolutionary AI model. This model is capable of guiding robotic devices to perform complex tasks they have not been explicitly or directly trained on in the past. This advanced capability is considered a massive step that has astonished researchers working at the company itself, opening unprecedented horizons in how machines interact with their surrounding environment and understand physical variables.
The concept of compositional generalization and skill repurposing
The new model, dubbed “π0.7”, showcases what is known in technical circles as “compositional generalization”. This capability lies in the possibility of combining a set of motor skills that a robot has learned in completely separate and different contexts, and using them to solve entirely new problems in the real world. The company describes these results as early signs of a fundamental and comprehensive shift in the ways machines learn—a shift that may emulate in its speed and scope the massive gains recently witnessed in the capabilities of large language models.
Instead of relying on the laborious process of gathering massive data for each new task individually, the “π0.7” model treats motor skills like linguistic vocabulary that can be recombined and reshaped in real time to form meaningful sentences. In one striking laboratory test, the model successfully operated an air fryer it had almost never dealt with during the training period, relying on recalling relevant manipulation and handling skills it had learned in previous contexts and experiments.
Transcending device limits and missing data
The model’s achievements did not stop there; in another more complex test, the model managed to control an advanced dual-arm industrial robot to fold clothes. What is surprising here is that the model did not possess any prior training data for this specific task on this type of device, yet it managed to match the success rates achieved by human experts when attempting to perform the same task for the first time using a teleoperation system. This achievement highlights an exceptional ability for rapid learning and adaptation to different devices without the need for continuous reprogramming.
Multimodal prompting framework
Research teams indicate that the magic key to unlocking these capabilities lies in an innovative framework relying on multimodal prompting. Instead of relying on simple and traditional text commands, the model receives detailed step-by-step linguistic instructions, alongside visual sub-goals generated via advanced visual models.
Furthermore, the model is provided with strategic metadata guiding it on whether priority should be given to speed or accuracy while executing the required task. This complex mixture of inputs allows the system to understand the general context of the task and make smart decisions in fractions of a second.
The turning point in the path of artificial intelligence
Sergey Levine, a professor at the University of California, Berkeley and one of the company’s co-founders, described this historical turning point by saying: “Once the model crosses that critical threshold where it moves from merely executing things for which precise data has been gathered, to the phase of recombining things in entirely innovative and novel ways, operational capabilities begin to rise at rates exceeding linear growth compared to the available volume of data.” This statement confirms that we are on the threshold of a new era where machines learn in a way resembling deductive reasoning.
Outperforming specialized models
Benchmark tests showed that the comprehensive and versatile “π0.7” model was able to match or even exceed the performance of previous models specialized in specific tasks that were previously developed by the company. This superiority was manifested in complex activities such as preparing coffee, folding clothes precisely, and assembling cardboard boxes. Despite these impressive results, the company is keen to exercise caution in its claims, preferring to describe these achievements as merely initial demonstrations highlighting future capabilities rather than a final product ready for immediate commercial deployment.
A rocket leap in financial valuation
The publication of this advanced scientific research comes just a few weeks after economic reports stated that Physical Intelligence is holding advanced discussions to raise new funding worth approximately $1 billion, which could raise the company’s total valuation to over $11 billion.
This new valuation, if completed, represents approximately a doubling of the previous valuation of $5.6 billion following a $600 million funding round conducted just a few months prior. This massive investment round is expected to involve the investment fund of prominent investor Peter Thiel, alongside venture capital partners and a number of previous financial backers. Despite the strong momentum the company enjoys thanks to its groundbreaking innovations, the funding deal is still in the early stages of discussion.
FAQ
What is the π0.7 model developed by Physical Intelligence?
It is an advanced artificial intelligence model designed to guide robotic devices to perform complex and diverse tasks they were not directly trained on, relying on the concept of compositional generalization to combine previous skills to solve entirely new problems.
How do smart robots manage to perform new tasks without prior training?
The model relies on a multimodal prompting framework; the system receives detailed linguistic instructions, visual sub-goals, and guiding strategies determining priorities for speed or accuracy, enabling it to adapt instantly to unfamiliar tasks.
What is the company’s currently expected financial valuation?
The company is holding advanced discussions to raise new funding worth $1 billion, which could raise its total valuation to over $11 billion, representing roughly double its previous financial valuation.
