Eka Robotics raises $150 million at $1.45 billion valuation for physical AI models

Eka Robotics completed a $150 million Series A round, bringing its post-money valuation to $1.45 billion, TechCrunch reported on October 20. The round was led by a16z and F-Prime. The company touts a manipulation architecture that includes force-sensing grippers and new deep learning models. According to TechCrunch, its robots can pick up about 1,500 objects per hour, weighing up to 1 kg.

The round, led by a16z and F-Prime, and the $1.45 billion valuation suggest that the combination of grasping performance and learning models is becoming a factor in how robotics companies are valued by capital markets. According to TechCrunch, the company says its robots can handle about 1,500 objects per hour, weighing up to 1 kg. Whether that approaches the throughput required in warehouses and manufacturing facilities will be a key point to assess.

The focus is not simply on Eka Robotics’ fundraising, but on how its valuation is tied to a manipulation architecture that combines force-sensing grippers with new deep learning models. What is new is that the company is selling grasping, control and learning as an integrated offering, rather than simply raising funds as a robotics startup. Meanwhile, the figures reported by TechCrunch — about 1,500 objects per hour, weighing up to 1 kg — can be read as an indication of throughput, rather than an impression based on a demo. In contrast, this time the basis for evaluation at the point of funding is not the visual appeal of a video, but “how objects are grasped and how quickly they can be handled.” In other words, the discussion appears to be shifting from the buzz around demos to the reproducibility of manipulation performance and the validity of the learning architecture. From an industry-structure perspective, the news shows that the manipulation platform combining grippers and models is itself becoming an investment target, alongside the quality of the robot hardware. If reports that Eka’s robots handled a variety of objects without retraining are accurate, the issue shifts from individual hardware improvements to how force-sensing inputs are connected to learning and how retries are accommodated after failures. What to watch next is the set of working conditions under which the rate of 1,500 objects per hour can be maintained, how broadly the range of object types and weights can be extended, and whether the performance can withstand sustained operation beyond research demos.

The article does not identify any business partners or customers in Japan. However, if the trend toward evaluating force-sensing grippers and learning models as an integrated system gains strength, the competitiveness of Japanese companies’ precision hands, force sensors and control systems is likely to come under scrutiny.