Eka Robotics, a Cambridge, Massachusetts startup building robots that handle objects by touch as well as sight, has raised a $150M Series A. Andreessen Horowitz (a16z) and F-Prime led the round, Eka announced on its blog on Tuesday. The round values the company at $1.45B post-money, Tim Fernholz reported for TechCrunch.
F-Prime is a venture fund linked to FMR, the parent company of Fidelity Investments. Obvious Ventures and E14 Fund, which led Eka’s seed round, invested again. MFV Partners, Bison Ventures, Glasswing Ventures, Samsung Next and several unnamed large companies also took part. Eka says the money will help it train its model on more tasks and speed up customer deployments.
Touch, not only sight
Many robot AI models connect what a robot sees and the instructions it gets to the actions it takes. Eka’s model adds force. It combines vision with feedback from force sensors in the robot’s hand, so the robot can feel a grip slipping and correct it.
“Vision is not the primary way in which you close this last millimeter problem. Vision tells you roughly where the object is, what the pose is, but the actual grabbing, the actual searching is happening through forces,” co-founder Pulkit Agrawal told TechCrunch.
Eka designed its own hand for this, with more force sensors than a standard gripper, TechCrunch reported. Agrawal said the hardware costs less than $1,000. It is rated for 5 million duty cycles, and the goal is 20 million, the level of standard industrial arms.
Others are working on robot touch too. Zurich’s Enact raised $30M this week to give robots a sense of touch. In China, robot-hand maker Linkerbot was seeking a $6bn valuation in May.
Practice in simulation
Much of the robotics industry teaches machines with recordings of people doing the task. Eka trains by trial and error in a physics simulator instead, using reinforcement learning, then moves the skills onto real hardware. It says it can generate thousands of hours of robot experience in minutes.
The founders met as graduate students at the University of California, Berkeley. Agrawal leads the Improbable AI Lab at MIT. Tuomas Haarnoja worked at Google DeepMind and co-created Soft Actor-Critic, a widely used reinforcement learning algorithm.
Other startups collect human data instead. Maxinsights raised $30M this week to turn household chores into robot training data.
1,500 picks an hour
Eka says its robots pick objects they were not trained on at 1,500 an hour, about two and a half times the human rate of around 600. The same robot can screw in a light bulb and insert wires. According to TechCrunch, it handles objects of up to 1 kg and succeeds more than 95% of the time. TNW has not verified those figures.
“Eka’s Vision-Force-Action model is the first approach I’ve seen that truly solves manipulation with the generality and throughput needed for deployments,” said Raghu Raghuram of a16z in Eka’s announcement.
Kris Hauser, a professor at the University of Illinois Urbana-Champaign, told TechCrunch the robot had a “wow factor” and adapted to many objects without retraining.
The factory test
Eka is running pilots with industrial customers in pick-and-place and wire handling, TechCrunch reported. It has not named them. Hauser said the real test comes on the production line.
“What’s going to be a big challenge is when they work with integrators on real assembly lines, and it needs to run without breaking for an entire year,” Hauser told TechCrunch.