Physical Intelligence, a robotics startup based in San Francisco, has recently released research that showcases its innovative model, π0.7, capable of directing robots to undertake tasks they were not specifically trained for—an unforeseen advancement even for the company’s own researchers. This marks a significant step toward creating a versatile robotic brain that can learn and adapt to new commands using plain language instruction.
The concept at the heart of this research is compositional generalization, which enables robots to combine skills learned from various contexts to solve unfamiliar problems, breaking away from traditional training methods that relied on memorising data for specific tasks. Co-founder Sergey Levine noted that this model could improve capabilities exponentially as it moves beyond standard training protocols.
One remarkable demonstration involved an air fryer, with π0.7 managing to understand and operate the appliance despite minimal direct training examples. It could even cook a sweet potato by following verbal instructions like a human mentor would provide to a new employee, highlighting the system’s potential for real-time adaptability in novel environments without needing additional data collection.
Researchers are conscious of the model’s current limitations. For instance, its ability to execute complex multi-step tasks autonomously still requires human guidance for effective operation. Physical Intelligence’s testing regime compared π0.7 with its prior specialized robotic models and found that the generalist model held its ground across various tasks such as making coffee and folding laundry.
The unexpected successes of π0.7 have notably surprised the researchers, illustrating its potential to generalise learning in ways akin to breakthroughs seen in large language models. Levine expressed that similar to the early, shocking moments of large language models producing unexpected results, this robotics model has led to remarkable demonstrations beyond initial expectations.
Though acknowledging potential scepticism regarding the tasks’ complexity, Levine argued that the true value lies in the system’s fundamental ability to generalise rather than performing flashy stunts. The research is still in its early stages, and while the findings show promising signs, Physical Intelligence remains cautious about commercial rollout timelines.
Having secured over $1 billion in funding and a valuation of approximately $5.6 billion, the startup is positioned for further growth, with talks of a new funding round that could potentially double its valuation. This financial backing stems partly from co-founder Lachy Groom’s reputation as a successful angel investor, enabling the company to attract substantial institutional interest despite its lack of immediate commercial plans.
Fanpage: TechArena.au
Watch more about AI – Artificial Intelligence
