Tutor Intelligence launches second-generation Cassie and Sonny warehouse robots

Sonny began as Tutor's dual-armed semi-humanoid robot and was introduced alongside the company's Ti0 4.5 billion-parameter Vision Language Action model. It has since moved from a research platform to a fully deployable robot capable of autonomous manipulation and mobility in warehouses. Cassie, Tutor's large-format single-arm robot, was expanded for mobility in warehouses and factories, adding pallet movement and greater physical intelligence. It looks like a giant arm atop a large box with wheels.

Co-founder and Chief Executive Josh Gruenstein told SiliconANGLE in an interview that it was important for the company to build general robots that are immediately practically useful. He said the company looked at an industry where humanoid robots are currently "sexy," but humans already do that fundamental work just fine. Cassie and Sonny represent two separate answers to the same question: what shape can a general robot take if it is designed around the work rather than around looking like a person?

Cassie provides muscle for moving heavy material in a broad, big box-laden pallet warehouse setting, while Sonny works amid shelves where arms and hands make sense for picking up small objects and moving them between bins. Gruenstein described Sonny as a co-design with Cassie of something that is human-like as a robot but not designed just to mimic a human, instead doing a class of work that humans do. Cassie can lift 50 pounds and tug thousands of pounds, which Gruenstein described as something of a "loch ness" of a robot moving between pallets doing work. The first generation of Cassie worked from a fixed position, but it is now mobile and can move across warehouse floors.

In December, Tutor raised $34 million in Series A funding. Since then, the company built what is essentially a "classroom" for its AI models to run across its Sonny robots in what it calls Data Factory 1: a 100-robot factory where Sonny robots learn. Gruenstein explained that robotics does not have an internet of training data to train AI models. AI models depend on vast amounts of data for pre-training; in robotics, that data is visual and kinetic telemetry that shows robots how to do tasks such as picking up objects, what objects look like, where to place them and how to move. Much of this can be done through training, or "tutoring," via experts who show the robots how it is done in real settings.

According to Gruenstein, Cassie does not need to reason about everything a pair of hands might encounter. Its physical vocabulary is comparatively small: identify an object, grasp it, move it and place it. Sonny's vocabulary is much larger, so Tutor built Ti0, the company's powerful AI model, and DF1 to attack the corresponding data problem. "We have this team of tutors staffed internationally that can kind of remote control a robot and teach it how to do a new task for the first time," Gruenstein said.

After that, the 100 robots in the classroom go through the motions to fine-tune behaviors, smooth out any kinks and prepare the robotic models for what might happen in the field. This is known as post-training. If an individual robot does a good job, that student's actions get reinforced in the AI model with a thumbs up, a positive reward; if it does poorly, those actions get a thumbs down. This helps the model learn how real-world behaviors interact with actual hardware. "We can get robots that recursively self-improve and are superhuman in their performance relative to people," Gruenstein said.