Rhoda's robot finds textured balls by touch, with its cameras blocked

In short: Rhoda AI published a tactile sensor and a version of its Direct Video-Action model that predicts future video and future touch. With wrist and head cameras blocked, a robot trained on one smooth ball and one textured ball then picked textured balls out of a box of seven, including two textures it had not trained on.

The sensor

• Rhoda says a gripper-mounted version responds to a feather (a few millinewtons) and to a dumbbell (over 40 newtons), and still feels the feather after a heavy load. Off-the-shelf optical and magnetic sensors it tested showed no clear response to the same feather.
• It reports normal and shear force, contact vibration sampled at about 500 Hz, and temperature, demonstrated with warm water and ice on a 20–40 °C scale.
• The same sensors go into a handheld Universal Manipulation Interface, so human demonstrations record contact force along with video and motion. Human demonstrations recorded this way are the training data for imitation learning.

The blindfolded task

The model takes a history of tactile "videos," predicts the next tactile signals, and an inverse-dynamics model turns those predictions into motion. Cameras on the wrist and head were covered. The robot had to search a box, tell a textured ball from a smooth one by contact, and adjust its grip.

Training used one smooth blue ball and one green textured ball. The same model then ran on four smooth balls and three textured balls. Two of the textured balls differed from the training ball in diameter, firmness and surface pattern. Rhoda says the model still identified and retrieved the textured balls, and shows an extended autonomous run. It does not report how many trials that took, or a success rate. Humanoids Daily notes that the video includes wrong-ball picks and re-scans, with some sections sped up to 6× and 12×.

Why it matters

Vision fails when the gripper hides the contact. Rhoda treats touch as another video stream inside a model built to predict pixels, for embodied AI that still has to act when the camera cannot see. The demo is a controlled box of balls, not a factory bin. Rhoda does not report reliability across repeated trials.