Rhoda AI builds its own tactile sensor and lets its video model steer a blindfolded robot by touch

Rhoda AI has taught a robot to find an object it cannot see. In a research post dated October 13, the Palo Alto company said it blocked a robot's head and wrist cameras, put a smooth ball and a textured ball in a box, and let its model guide the gripper by touch alone until it found, grasped and removed the textured one. Trained on just those two balls, the same model then retrieved all three textured balls from a box of seven, two of which differed from the training ball in diameter, firmness and texture, according to the company.

The demonstration rests on two things Rhoda built itself: a tactile sensor and a version of its Direct Video-Action (DVA) model that predicts touch as well as images. Both matter more than the ball trick, because they show how the company plans to spend the $450 million it raised seven months ago.

A sensor that feels a feather and carries 40 newtons

Rhoda said its sensor, packaged here to fit a robot gripper, registers a feather's touch of a few millinewtons, which the off-the-shelf optical and magnetic sensors it tested did not distinguish, and still measures loads above 40 N from a dumbbell. It separates normal from shear forces, samples contact vibrations at about 500 Hz, which is enough to tell a spiral binding from a zipper, picks up a wrist pulse and reads temperature across its surface. The company did not disclose the sensing principle, price or whether it will sell the sensor, describing it only as compact, low-cost and co-designed with calibration algorithms that reduce drift and hysteresis.

Rhoda also fitted the sensors to a handheld Universal Manipulation Interface, the gripper people use to record demonstrations without a robot, so human demonstrations now carry the forces the demonstrator applied.

How a video model learns to feel

DVA is Rhoda's bet that a web-scale video model can act as the robot policy: it predicts what the next moments should look like, and a separate inverse dynamics model turns that prediction into joint commands. The new multimodal version takes a history of tactile 'videos' as input and generates future tactile frames; the inverse dynamics model converts them into actions, and fresh sensor readings close the loop. When vision is missing, touch alone can drive the search.