Dexterity is a battery problem before it's an AI problem — and assistance is a learning problem before it's a mechanical one. US12420401B2, granted to North Carolina State University in September 2025, covers "Multimodal end-to-end learning for continuous control of exoskeletons for versatile activities."

The B25J filing tells a different story than the keynote. Classified under B25J 9/0006, B25J 9/1615 and B25J 9/163, the patent replaces the usual exoskeleton design — a set of hand-tuned controllers, one per activity like walking, climbing, lifting — with a single learned, end-to-end controller that adapts continuously to what the wearer is actually doing.

“Various examples are provided related to continuous control of exoskeletons. In one example, a method includes obtaining IMU sensor signals associated with an exoskeleton attached to a limb of a subject; generating an exoskeleton control signal in response to the IMU sensor signals, the exoskeleton…”— U.S. Patent No. 12,420,401 source

Here is why 'versatile activities' is the hard, important phrase. Earlier exoskeletons worked because they assumed a known activity; switch from walking to stairs and a pre-scripted controller fights the wearer. A learned controller that reads multimodal signals — motion, force, intent cues — and adapts on the fly is what lets one device assist across the messy variety of real human movement.

The exoskeleton is the human-in-the-loop cousin of the humanoid. Both face the same core problem: controlling a legged, dynamic system in contact with an unpredictable world. The exoskeleton adds a partner — the wearer — whose own intentions the controller must infer and assist rather than override. That makes intent-reading central in a way a standalone humanoid can dodge.

The honest limit is safety and trust. A learned controller strapped to a human body cannot be allowed to explore dangerously or behave unpredictably, and the gap between a learned policy that works in a lab and one a person trusts to support their weight on a staircase is exactly where these devices live or die.

For readers auditing embodied AI, the exoskeleton work is a useful tell. The same end-to-end learning trend reshaping humanoid control is reshaping wearable assistance — and because the exoskeleton has a human in the loop, it surfaces the intent and trust problems that standalone humanoids will eventually have to confront too.