Touch has moved from research labs onto humanoid product specification sheets. Figure’s Figure 03 hands carry a tactile sensor in each fingertip, and Boston Dynamics’ Atlas specification sheet, revised October 15, 2026, lists “Tactile fingers and palm” under sensing. What has moved more slowly is public evidence of what that sensing does for a robot’s work.

What Figure has shown

Figure’s Helix 02 announcement in January 2026 is the clearest public example of touch feeding a learned control policy. Figure says its fingertip sensors “detect forces as small as three grams,” enough to feel a paperclip, and that Helix 02 was “the first time we’ve demonstrated neural network policies that depend on these modalities,” meaning the tactile sensors and the palm cameras added in Figure 03.

The tasks Figure picked to show it depend on contact: unscrewing a bottle cap without crushing the bottle, picking one pill out of a medicine organizer, pushing exactly 5 ml from a syringe, and pulling small metal parts out of a cluttered box. The post presents those four tasks as videos, without success rates.

What Boston Dynamics lists

The Atlas sheet describes a 56-degree-of-freedom robot with tactile fingers and palm and a 360-degree camera view. It does not say what the tactile sensors measure, at what resolution, or how Atlas’s software uses the signal.

The research tools

Much of the open tactile research comes from Meta’s FAIR lab, which released three pieces on October 31, 2024. Sparsh is a family of touch-perception models pretrained without labels on more than 460,000 tactile images. Digit 360 is a fingertip-shaped sensor with more than 18 sensing features. Digit Plexus is a hardware and software platform for connecting several tactile sensors on one robot hand. Meta named GelSight to manufacture and distribute Digit 360 and Wonik Robotics to produce an Allegro Hand with tactile sensors. These are research platforms, not parts of a humanoid product.

What touch is for

A camera loses sight of the contact once fingers close around an object. Touch reports what is happening there: how hard the grip is and whether the object is starting to move. Figure describes the payoff as “contact-aware, force-modulated grasping.”

The evidence that would settle how much touch adds is a comparison on the same robot and task with the tactile input switched on and off, reported as drops, damaged items and completion rates. None of the sources here reports that comparison for a humanoid.