Dexterity is a battery problem before it's an AI problem — unless you let the gripper's body do the thinking. US11565406B2 ("Multi-tentacular soft robotic grippers") and US11685058B2 ("Soft robotic tentacle gripper"), both granted to Mitsubishi Electric Research Laboratories in 2023, take that route.

The B25J filing tells a different story than the keynote. Classified under B25J 15/12, B25J 9/1612 and B25J 15/10, the soft tentacle gripper grasps not by computing finger positions precisely but by wrapping compliant tentacles around an object until they conform. The mechanism does the adapting that, in a rigid hand, the controller would have to compute.

“A gripper system having tentacles including a control system configured to receive operator data and sensor data.”— U.S. Patent No. 11,565,406 source

Read past the abstract and the control architecture is more specific than "let the body conform." Claim 1 describes a centralized control system that, given a target object's shape and pose from sensor data, looks up a matching object configuration in a stored database, and with it an associated set of grips. From there it retrieves stored sets of commands tied to that configuration. The branching logic is the interesting part: if the operator data includes a set of pickup actions, the system matches those against the stored pickup actions for the configuration and selects a first set of commands; if the operator supplies no pickup actions, it falls back to a second set. Either way, the output is "a sequence of control signals that cause motors for each tentacle… to apply a sequence of tensions to transmission systems," moving the object. The intelligence is in the lookup-and-match, not in solving a grasp from first principles each time.

The dependent claims fill in the hardware doing the conforming. Each tentacle, per claims 7–9, is built from lower and upper members joined by a connector with a center thru-hole and tangential transfer channels; guide discs spaced along each member define cable pathways, and cables run from motors through a "controllable palm baseplate" out to distal guide rings. This is a cable-driven, tendon-actuated continuum finger — the tension sequence the controller emits is literally pulling cables that bend the tentacle. Claim 5 makes that explicit: the command sets are "a sequence of predetermined torques to be applied to… joints or flexural joints along with corresponding tensions," each transmission running from a motor at one end to a joint at the other.

Sensing closes the loop. Claim 10 puts an end-tip sensor on each tentacle — an inertial unit, a MEMS device, an accelerometer, or an electromagnetic tracker — feeding a "tentacle state function." Claim 11 adds tactile signals embedded in the tentacle's outer surface: measured pressure, moisture, shear force, and torque. Claims 12–13 add distal-joint shape or linear-displacement sensors (an LVDT, a Hall-effect sensor and magnet, a slide potentiometer) that infer how far each cable has moved and therefore how the joint has bent. That matters because a continuum body's tip position is not given by joint encoders the way a rigid arm's is; it has to be reconstructed from cable travel and shape sensing.

Here is the philosophical split this patent dramatizes. The rigid-hand camp — the one most humanoids belong to — puts intelligence in control: precise joints, precise sensing, precise planning. The soft-robotics camp puts intelligence in mechanics: a body so compliant that a crude command produces a good grasp because the material conforms to whatever it meets. The grip taxonomy in claims 2–3 shows the design has not abandoned structure, though — it distinguishes grip styles (internal vs. external coordinated grasps, built on kinematic constraints and pinching or clamping friction) and grip modes where one tentacle braces against another. The compliance is organized, not random.

The honest trade is precision for robustness. A soft tentacle gripper will happily grab an irregular, fragile or unfamiliar object that would defeat a rigid hand's grasp planner — but it cannot perform fine, positioned manipulation the way articulated fingers can. You buy forgiveness and lose finesse. The patent's failure-handling logic concedes as much: claim 6 has the controller confirm, after a grasp, that a single target object is held and that its position and orientation meet predetermined criteria, and claim 17 (and the recovery logic in claims 23–24) describes generating an "alternate set of command actions" — returning the object to its start position or selecting a subsidiary command sequence — when the grasp does not land where it should. A conforming gripper still needs a retry policy, because conforming is not the same as confirming.

One more grounded detail worth flagging: claim 19 trains those command sets by "sensing a motion having human like characteristics from a training operator wearing a teaching glove with sensors," then converting the signals into tentacle commands. The system's vocabulary of grasps is partly demonstrated by a human, not hand-coded — the same teaching-by-demonstration idea that recurs across the modern manipulation literature, here applied to a tendon-driven soft hand.

This is not a fringe idea; it is a serious counter-bet to the dexterous-humanoid-hand orthodoxy. For a lot of real picking — produce, soft goods, mixed bins — "conform and hold" beats "compute and pinch," and the soft-gripper patents are the IP behind that argument.

For readers auditing manipulation, the soft-gripper case widens the lens. The question is not only how many fingers and how many degrees of freedom, but whether the design puts its intelligence in software or in the gripper's own compliant body. Both can be right; the patent reveals which bet the engineer made — and Mitsubishi's bet, read claim by claim, is a tendon-driven tentacle that senses its own shape, looks up a grasp rather than solving one, and keeps a recovery routine on hand for when conformity is not enough.