The arm in the glass case
In Terminator 2, the future fits inside a glass case. At Cyberdyne, Miles Dyson studies a metal arm and a processor recovered from the first Terminator. The machine is gone. Enough of it remains to give the engineers somewhere to begin.1
Thirty-five years later, a humanoid-hand patent can produce a flicker of recognition. There are the metal fingers. There is the crowded forearm. Look closer and the drawing becomes a collection of decisions: where to put the drives, how to turn a wrist without disturbing the grip, and how to make a finger bend.23
The Tesla side of this comparison centers on five newer international filings, submitted in October 2025 and published in April 2026. Together they cover the forearm, wrist, cable routing, flexible joints and the grasping appendage. They provide a much richer picture than the older floating-cable-end finger. Each remains a disclosed design; the records do not establish that every embodiment appears together in one production Optimus hand.30424324445
Figure supplies two useful comparisons: forearm-driven tendons and removable fingers with motors inside them. Follow the force from the drive to the object. The interesting differences appear along the way.365
When intelligence reaches for something
The prospect of recursive self-improvement gives the old movie image a new charge: an AI system improves its ability to make further improvements. There are already bounded experiments in that direction. Sakana AI’s Darwin Gödel Machine rewrites its agent code and evaluates new versions on programming tasks. That is evidence of a specific software experiment, with a long distance still between it and an open-ended intelligence explosion.7
On the hardware side, Figure’s January 2026 Helix 02 announcement shows a robot working through a four-minute dishwasher task and manipulating small objects. Figure describes the demonstrations as autonomous. They offer a concrete glimpse of learned control reaching through a physical hand; they are company demonstrations, without an established mapping to the patent embodiments examined here.8
Software self-improvement can happen without a humanoid body. But if more capable AI is to assemble equipment, handle tools or conduct physical experiments, the machinery has to deliver. A better plan still needs a grip that works. These patents let us examine that part of the problem in detail.
Same destination for force, different drive packages
Both forearm-driven designs move machinery out of the fingers. The meaningful difference is how they package the outputs that must pass through the wrist. Tesla’s newer forearm filing arranges longitudinal linear actuators around larger wrist actuators, with a separate rotary stage for roll. Figure describes rotary tendon drives, including cycloidal transmissions, at different distances from the wrist.23
Tesla’s arrangement has to accommodate actuator strokes along the arm and collect their pulling outputs at its narrow end. Figure’s arrangement has to accommodate the rotary drives and the tendon paths leaving their distributed positions. Both buy space in the hand by accepting a transmission path through a moving wrist; changing the drive package changes the space that path must negotiate.23
The practical comparison is packaging efficiency at a required force and stroke. Tesla attributes assembly benefits to separating the rotary stage from the linear bank. For both layouts, usable comparisons would measure access for assembly, heat under repeated gripping and the work needed to replace one drive. The drawings identify those tradeoffs without measuring their outcome.23
Which route keeps the grip steady as the wrist turns?
Hold a cup and turn the wrist. In both tendon-driven hands, the finger still needs a controlled pull while joints around the tendon path move. If the route changes length unintentionally, wrist motion can disturb the grasp. That is the same problem for both designs.342
Figure addresses it through the tendon-departure positions, routing curvature and, in some embodiments, a carpal tunnel offset from the pitch axis. Tesla’s newer hand filing reorganizes the bundle itself: a lateral cable stack on one side of the wrist becomes a vertical stack on the other, with supports arranging the transition around the two rotation axes.342
The distinction is between where the tendons enter and curve through the joint, and how their arrangement changes across it. Both are geometric ways to limit interference between wrist and finger motion. Neither passage supplies measured grip stability. A direct test would hold commanded finger tension constant, sweep the same wrist motion and measure the unwanted change in fingertip force.342
The wrist support competes with the cable corridor
Routing cannot be compared separately from the mechanism supporting the wrist. Figure’s carpal-tunnel arrangement reserves a path through the joint for the finger tendons. Tesla’s newer wrist filing makes space beneath a cantilevered universal-joint arrangement and uses a notch to accommodate the structure during extension. Both must fit support and tendon clearance into the same volume.343
Tesla also describes two drives acting together through side links: coordinated movement produces pitch and differential movement produces yaw. That is a specific actuation mechanism. Figure’s cited routing passage is not an equivalent complete wrist-drive specification, so it cannot support a torque or range-of-motion ranking against Tesla’s linkage.433
The comparison we can make is the competition for space. The useful next measurement is how much wrist motion and load capacity each assembly retains while keeping the tendon corridor clear. Calling one a better wrist without that shared test would skip the constraint that connects them.
A remote tendon drive or a replaceable motorized finger?
Figure’s modular branch changes the comparison more fundamentally than its forearm layout does. Against Tesla’s remote drives and cable corridor, it places the motor and transmission inside a removable finger. US20250205908A1 describes four substantially identical removable fingers; published claim 10 specifies that each lacks a mechanical cable for actuating its components.6242
The related issued finger patent shows a motor turning a worm drive, which engages a perpendicular worm wheel and moves a linkage. Independent claim 16 combines removable motorized fingers with a member that biases part of the mechanism toward its uncurled position. The drive no longer needs to transmit finger actuation mechanically through the wrist.5
That is a concrete exchange: less remote actuation routing for more motor and transmission hardware at the hand. A repair can be organized around replacing a digit, while the remote-drive approach involves the tendon path back into the arm. The modular choice also puts heat, electrical connections and mechanism volume into the replaceable part. Replacement time, recalibration and repeated-grip temperature would show whether the exchange pays off.652
Figure’s motorized and forearm-driven disclosures are alternative branches. Combining their benefits into a single imagined Figure hand would make the comparison meaningless.
Rolling contact or an internal geared linkage?
At the finger joint, compare how motion is constrained. Tesla’s newer flexible-joint filing has curved surfaces roll against one another, so the rotational axis moves with the contact. A flexible composite member connects the two sides. Figure’s issued motorized finger instead organizes a worm wheel and a proximal link around a shared axis while allowing them to rotate independently.445
Tesla’s connection must bend readily while resisting unwanted stretch. Its dependent claims describe reinforced layers and, in some versions, an electrical harness between layers. Figure’s mechanism has to package the geared drive and linked motion inside the digit. These are different mechanical interfaces to keep accurate under repeated loading.445
That difference suggests a direct durability comparison: after the same gripping cycles, measure lost motion, stiffness and force repeatability. For the Tesla embodiment, inspect the reinforced flexible connection and rolling surfaces; for Figure’s, inspect the gearing and link interfaces. The records identify different places for wear to matter. They do not report a matched fatigue test.
Specialize the pinky or share the mechanism?
Give both hands the same awkward object. Tesla’s newer appendage filing gives the pinky an opposition joint, letting the edge of the hand turn inward to help wrap around it. Figure’s cited forearm embodiment couples ring- and little-finger motion; its modular branch instead emphasizes repeated, substantially identical removable fingers. The contrast is how much specialized motion and hardware to allocate to that edge of the grasp.4536
Tesla also separates outer-joint curling from paired paths at the finger base: unequal tension spreads the finger sideways, while pulling both can flex the base. That gives the disclosed mechanism several ways to shape contact. Figure’s coupling shares movement between digits, while repeated modules offer a common part design. Those choices trade different kinds of motion control against drive sharing or part commonality.4536
More specialized motion only helps if the task uses it. A useful comparison would measure how each arrangement stabilizes the same shapes, and which grasps require independent ring/pinky movement. Actuator counts alone cannot answer that question.
Same contact, two different sensing questions
Press the same object into both hands. Figure’s tactile grant asks how force passes through the covering material into a strain-gauge assembly. Tesla’s newer appendage disclosure describes where sensors sit: along finger segments, opposition members and the palm, with wiring through the joints. One record details the force-transfer layer; the other details sensing coverage and routing.945
Figure’s issued independent claim specifies interior and exterior material regions with different compression/deflection ratios, alongside the sensor, electronics and robot. That material response can influence what reaches the sensing structure. Tesla’s distributed locations address which working surfaces can supply contact information. These are complementary requirements, so the present evidence cannot support a claim that one company has tactile sensing and the other lacks it.945
The direct comparison would apply the same contact forces and incipient slips across the working surfaces, then repeat after wear. We would need an equally detailed Tesla tactile claim and hardware review to compare its force-transfer construction with Figure’s. The gap belongs in the evidence, rather than being filled with a presumed design difference.
A useful hand has to do it again
Picture a hand lifting one cup. Then picture it doing that job all day, at changing wrist angles and with worn contact surfaces. The comparison becomes a set of demanding tests.
For the tendon-driven hands, measure whether wrist motion disturbs finger force. For Tesla’s rolling joints, measure stiffness, hysteresis and fatigue in the flexible connections. For Figure’s motorized digits, measure heat, connection reliability and the time needed to replace and recalibrate a module. Then test how both hands maintain contact information as their surfaces wear. These are questions suggested by the disclosed mechanisms, not results reported here.4244369
The file histories add a different kind of evidence: how specific claim language fares against earlier work. Figure’s changing tendon-route examination and Tesla’s older cable-end argument are preserved below. They help explain what the applicants sought to protect; the newer mechanisms deserve the center of this engineering comparison.131819
The future has to fit inside the hand
The arm in the glass case arrived as a finished mystery. The newer Tesla filings let us work inward: from a crowded forearm, through a wrist that reorganizes its cables, to rolling finger joints. Figure’s portfolio puts different drive layouts and removable motorized digits beside that picture.2424435
The interesting contest is how those choices support useful work. Can the hand maintain its grasp as the wrist turns? Can it adapt contact around an awkward object? Can a worn part be replaced without rebuilding the whole mechanism?
More capable AI would make those questions more consequential. A system that can devise a better experiment still needs dependable machinery to carry it out. These patents establish no timetable for that future and no performance winner. They reveal enough of the machinery to make the questions specific.
A useful robot hand has to grasp, feel and keep working. That ambition is becoming a set of parts we can inspect.
Evidence notes, claim details and open questions
Figure’s promising route meets an older hand
For a moment, Figure appeared to have found a way forward. In December 2025, the examiner identified a dependent tendon-routing claim as allowable if rewritten with all its inherited limitations. Figure later brought the route into its independent claims: below the knuckle, above the proximal phalanx, below the medial phalanx.1112
Then the search found another hand. The July 2026 action introduced Wang CN115070744A, indexed to Zhejiang University of Technology, against the routing limitation. All 28 pending claims were rejected. A geometry that had looked promising now had to be distinguished from a newly cited reference.131415
That turn earns its place in the engineering story. The route is both a way of moving a finger and a proposed distinction from earlier designs. The examination asks whether the claimed combination actually supplies that distinction. Figure’s granted motorized-finger and tactile claims remain separate outcomes; this forearm contest is still open.1359
Earlier Tesla hand: the floating cable end and its examination
The older WO2024073138A1 gives an enlarged cable end room to move inside a region of the distal finger member, with the stated aim of accommodating movement and limiting cable damage. This is background to the newer hand comparison, not evidence of the newer rolling joint’s construction.4
Tesla’s older hand history also turns on a small piece of geometry. The first international opinion challenged its claims using DE102020207037B4, whose publication names KUKA Deutschland GmbH as owner, alone or with another reference. The opinion had already considered dependent claims containing the floating-end feature.1617
Tesla’s response asked the authority to look again at where the cable terminated. It amended the independent claims and argued that KUKA’s cable end sat outside the fingertip, while Tesla’s claimed end could move inside a region of the distal member. The later international report accepted the distinction and gave favorable findings for the amended claims.1819
The location mattered. A phrase such as “floating cable end” was not enough to explain the difference; the argument depended on the physical relationship between the cable and the fingertip. The favorable report is an international examination result, not a US grant. The unresolved US claim-record question is preserved in the evidence notes.2010
Why the actuator counts cannot settle the comparison
The numbers look easy to compare until their definitions come into view. Tesla’s first forearm embodiment has 25 linear actuators—23 for the hand and two for the wrist—with a separate rotary actuator for roll. Figure’s cited forearm embodiment has twelve actuators including twist. Those are counts in particular embodiments, with different drive arrangements and allocations.23
Tesla’s separate whole-appendage embodiment specifies seventeen finger actuators for twenty-two finger degrees of freedom. That count belongs to that embodiment; it must not be merged with the forearm publication’s twenty-five linear actuators.45
A degree of freedom is a movement coordinate; an actuator is a drive. The relationship depends on the transmission and on whether joints move independently or are coupled. Figure, for example, describes coupling between ring- and little-finger motion. Tesla’s forearm claims also use different minimum actuator thresholds from its first illustrated example.32
A useful engineering comparison would follow each drive to the joints it controls, then measure force, speed and independence under the same task. Dividing one published actuator count by the other would skip that work. These records support an architecture comparison; they do not supply a dexterity ranking.
The earlier hands behind both examinations
The two records repeatedly lead back to earlier institutional robotics work. Figure’s July action uses Ihrke US20110071671A1 on wrist structure and US20110071673A1 on actuator packaging. The retrieved US assignment chains record interests in GM Global Technology Operations LLC and the United States represented by NASA. This attribution comes from assignment records, not from an inventor’s name.1321222324
Tesla’s newer forearm search cites Reiland US20120194120A1, whose retrieved assignment data likewise records GM/NASA transfers. The GM/NASA publication used against Tesla’s older hand is a different record, US20130193704A1: its front page names those assignees, but the retrieved child-application assignments show a security interest and release rather than a complete ownership chain.25262728
There are also two different Wangs. Figure’s cited CN115070744A is indexed to Zhejiang University of Technology. Tesla’s forearm report cites Wang Pengbo’s CN106737789A, whose publication names Jiangsu Jingang Culture & Technology Group as applicant. These are distinct references and attributions. The foreign records here have not received a complete assignment-chain audit or full technical translation; their role in the rejections is reported from the examiners’ mappings.1415292513
Shared prior art helps locate the engineering lineage. It does not show that one company copied the other, or that a citation defeats every claim in a family. Tesla’s forearm search, for example, covered claims 1–15 only after additional search fees were not timely paid. It supplies no merits result for the other two invention groups, claims 16–19 and claim 20.25
Read the family trees as relationships, not a race
The native patent-platform trees include the PCT applications that make Tesla’s dates intelligible. The five newer Tesla chains show October 10, 2024 provisional filings, October 9, 2025 PCT filings and November 25, 2025 US national-stage entries. Looking only at the last date would skip the intervening international filings.30
Figure’s broader exported family contains 114 nodes, including four PCT applications, with the forearm, finger and tactile reading points highlighted. Its size does not measure hand-patent strength, and continuity lines alone do not establish that every claim is supported by the earliest provisional. The trees are navigation aids for inspecting each branch’s documents and relationships.31
Both original investigations remain linked above, and both companies’ original SVG exports are available here. Keeping the documents attached to their own branch avoids turning multiple embodiments and procedural histories into a fictional single hand.
Claim wording and the unresolved US record
Figure’s modular branch changes the location of the drive. US20250205908A1 describes four substantially identical, removable fingers. Published claim 10 combines their arrangement on a frame with the absence, in each finger, of a mechanical cable for actuating its components. A June 26, 2025 amendment preserves a modular, cable-free configuration while revising the claim wording. The application remains pending in the platform’s current record.63233
The claim structure matters. A third, intermediate region appears in dependent claim 2. The three-flexure-arm sensing bridge appears in dependent claim 8. Neither is a universal requirement of independent claim 1. The original prosecution numbering has been reconciled with the issued claims: former claim 16 became claim 1, and former claim 17 became claim 2.93435
Figure’s finger allowance specifically identifies the independently rotating proximal link within the claimed combination.36
Figure’s forearm history shows why the drawing alone is only half the story. In December 2025, the examiner identified then-dependent claim 29 as allowable if rewritten with its inherited limitations. Figure’s attempted amendment after final rejection was not entered; it then requested continued examination and filed amended claims on March 4, 2026.113712
Tesla’s older hand took another path. Its February 2024 international written opinion challenged the published claims using DE102020207037B4, which names KUKA Deutschland GmbH as owner, alone or with the GM/NASA publication US20130193704A1. The challenged claims already included dependent floating-end limitations. The later favorable result cannot be explained simply as adding a feature the search had never considered.1641728
That is a documented change in the international examination, tied to an amended claim set and a specific argument. It is not a US allowance: international preliminary examination does not bind national offices. The older US file also contains a twenty-claim submission and later worksheet that have not yet been fully reconciled with the eighteen-claim international annex. The operative US claim set remains an open check.2038394010
Research scope and open questions
This comparison now centers Tesla’s five newer publications: WO2026080687A1 (hand and cable routing), WO2026080690A1 (wrist), WO2026080691A1 (forearm), WO2026080693A1 (flexible joint) and WO2026080701A1 (appendage and grasp). Their 2025 PCT filing dates and April 2026 publication dates distinguish them from the older hand history kept in these notes.42432444530
Four original newer publication PDFs remain missing from our archive. Their available description and claim transcriptions have been reviewed and are linked as publication text; they are not labeled as self-hosted originals. WIPO lists the documents, but direct and browser download attempts did not yield usable local files on September 13. The new overview image comes from the original forearm publication that is archived. Original drawings and wording in the other four publications still require visual cross-checking.424344452
A fresh patent-platform check covers the five newer US national-stage records and their IFW indexes. All five show pre-examination processing in the returned data. The available one-page SPEC placeholders do not constitute complete specifications. Separately, the earlier eleven-application check supports the Figure and older Tesla records. Neither check guarantees current upstream synchronization.4633
Tesla’s appendage description supplies a sensing layout. We have not completed an equivalent Tesla tactile claim-history review against Figure’s issued material-and-sensor combination. Published architecture, claim wording, prosecution results and demonstrated products remain separate evidence categories.109
Outstanding checks include the four newer publication PDFs and separate written opinions, the older Tesla US claim-set discrepancy, Figure’s next forearm response, fuller foreign-prior-art review and the Figure distance-ratio discrepancy. Downloaded documents and available OCR extend beyond the records substantively reviewed.1041
Sakana’s software experiment and Figure’s Helix announcement provide context for why capable physical systems matter. They do not establish imminent RSI, self-manufacturing robots or a mapping from a demonstration to these particular patent embodiments.78