Former Optimus Engineer Downloaded Tesla Hand Files Before Launching Rival

The Allegation: Tesla's Complaint

Tesla filed a 20-page complaint in federal court on June 11, 2025, accusing a former Optimus engineer of walking out with the company's most sensitive robotic-hand research and launching a competing startup within days. The case, Tesla, Inc. v. Proception, Inc. et al, Docket No. 5:25-cv-04963, landed in the U.S. District Court for the Northern District of California and named Zhongjie "Jay" Li and his company Proception, Inc. as defendants.

The lawsuit crystallizes a broader collision: the race to human-level robot dexterity has made the humanoid hand the most contested intellectual property in robotics, and trade-secret law is becoming the primary weapon for policing the boundary between an engineer's expertise and an employer's crown jewels. Tesla's complaint alleges a "calculated effort to exploit Tesla's investments, insights, and intellectual property" tied to Optimus, investments intended for factory automation and broader industrial use. In its Q1 2026 update, Tesla said first-generation Optimus production lines were being installed at Fremont in anticipation of volume production.

Li worked on Tesla's Optimus program from August 2022 until September 2024. In the weeks before his departure, the complaint says, he downloaded confidential design files, actuator specifications, sensor data, source code, and prototype videos onto two personal smartphones. Six days after his last day, Proception was incorporated. Within five months, the startup publicly claimed it had "successfully built" advanced humanoid robotic hands that Tesla says bear a striking resemblance to the designs Li worked on.

Tesla's counsel wrote that defendants "misappropriated Tesla's most sensitive materials, sidestepped the laborious process of development, and launched a company based not on original discovery, but on stolen work." Attorney Josh Krevitt added in the filing: "Innovation must be earned, not stolen. If misconduct like this goes unchecked, it risks incentivizing theft over creation and undermining the very ecosystem that drives progress."

Tesla is pursuing actual and punitive monetary damages, an injunction barring Li and Proception from using or disclosing the alleged trade secrets, and a jury trial. Li, who lists himself as Proception's founder and CEO, did not respond to requests for comment at the time of filing. No attorneys for Li or Proception appeared in initial court filings. The company, backed by Y Combinator and based in Palo Alto, had not publicly responded to the allegations as of the complaint date.

The lawsuit follows a pattern: Tesla has repeatedly pursued IP theft cases, including actions over Autopilot source code and battery technology. But this one targets the humanoid hand — the component that determines whether a robot becomes useful or remains a stage demo. The court's response and Proception's defense would shape what engineers can carry between companies in the race to dexterity.

Why the Hand Became the Battleground

Robot hands sit at the center of the next AI race because humanoid robots can't be useful in the real world if they can't pick up, rotate, grip, and handle objects properly. A robot that can talk but can't hold a tool is still mostly theater. Kevin Lynch, director of Northwestern University's Center for Robotics and Biosystems, told the Wall Street Journal last year that his team believes it will be a decade until hands are "functional and useful and able to do some of the things that humans do." Elon Musk has said robot hands are one of the biggest engineering problems yet to be solved. That difficulty is exactly why the IP around them has become a battleground.

ProHand 1.0, Proception's flagship product, is a 22-degree-of-freedom, tendon-driven robotic hand with multiple joints per finger enabling what the company calls "a wide range of dexterous motions." The tendon-driven actuation mimics human anatomy: motors pull cables that move the fingers while keeping the fingers themselves lightweight and compact. That mechanical choice matters. It pushes mass toward the palm or forearm, reducing inertia at the fingertips — critical for speed, precision, and impact tolerance. The hand also integrates skin-like sensors across its surface. Those sensors detect contact and support grip control during manipulation, closing the loop between touch and motion.

The hard-to-replicate IP isn't any single component. It's the integration. Force estimation errors below 0.2 N (as demonstrated in the DexHand 021 research platform) require sensor fusion calibrated against tendon tension, joint position, and motor current. Compared to PID control, advanced torque control in multi-object grasping reduces joint torques by roughly one-third while preventing overload during collisions. That level of coordination across mechanical design, sensor placement, firmware, and control algorithms is what takes years to develop. Tesla's complaint specifically targets "confidential information linked to advanced robotic hand sensors": the sensor layer and its integration with the control stack.

Proception's ProGlove adds another dimension. It uses the same sensor skin that encloses ProHand as a wearable data collection system for human hands. Human testers wear the glove (and a headset) to capture manipulation data without requiring a robot in the loop. Most robot manipulation data today comes from teleoperation, which doesn't scale easily because collection is constrained by the number of robots available and usually happens inside labs; it also loses important human interaction signals because the teleoperator receives no tactile feedback from the objects the robot touches. Proception's bet: hardware close enough to the human hand that human data becomes directly usable for training. Li told TechCrunch the combination of "highly dexterous hardware plus highly scalable data" is the key to solving dexterous manipulation faster than the decade timeline Lynch projected.

That hardware-data flywheel is what Tesla sought to protect. The complaint alleges Li downloaded sensitive files related to robotic hand sensors onto personal devices in the weeks before his September 2024 departure. Six days later, it was incorporated and within five months was demoing a five-fingered ProHand that Tesla calls a dead ringer for its design. Whether the resemblance stems from general engineering convergence or specific trade-secret transfer is what the court weighed.

How the Court Ruled

The case moved fast by federal-court standards. Tesla filed its complaint on June 11, 2025, before Judge Susan Van Keulen, alleging violations of the Defend Trade Secrets Act and seeking an injunction to stop Proception from using what Tesla claimed was its proprietary robotic-hand technology. Within a month, the court granted Tesla's motion for expedited discovery in part — allowing targeted forensic examination of Li's devices and Proception's systems, while denying broader requests that would have effectively halted the startup's engineering work. That early ruling, signed July 8, 2025, signaled the court would not treat the case as a routine trade-secrets dispute but would test Tesla's evidence against a functioning product.

The pivotal moment arrived November 14, 2025. Judge Van Keulen denied Tesla's motion for a preliminary injunction. The order, issued after sealed briefing and oral argument, meant Proception could continue developing and shipping its ProHand while the case proceeded. Courts deny preliminary injunctions when the movant fails to show a likelihood of success on the merits or irreparable harm that money damages cannot fix. The denial did not end the lawsuit, but it removed the immediate existential threat to Proception's operations. Tesla's legal team, led by Krevitt, continued pressing claims through discovery; Proception, represented by counsel from Bergeson, LLP, maintained that its tendon-driven architecture and tactile-sensor integration were developed independently after Li's departure.

Proception's public posture remained restrained. In a June 2026 interview with TechCrunch, Li characterized the litigation as a "resilience test," his term for the pressure of building a hardware startup while defending against a well-capitalized plaintiff. He did not detail the defense strategy. The litigation ended not with a verdict but with a stipulated dismissal with prejudice. On May 29, 2026, Tesla filed a stipulation proposing dismissal of all claims; Judge Van Keulen granted it on June 2, 2026, terminating the case. The settlement terms remain confidential. What is known: Proception closed an $11 million seed round led by First Round Capital, with participation from Y Combinator and BoxGroup, announced June 29, 2026, weeks after the dismissal. World IP Review noted the resolution "clears legal overhang for startup but highlights fierce competition for talent and trade secrets across the booming robotics industry."

For Tesla, the settlement closes a chapter in a pattern of aggressive IP enforcement against former employees who join or launch robotics ventures. For Proception, it converts a potential company-ending dispute into a capitalized runway. The company says it is now shipping its first production batch of ProHand units to research partners and opening wider orders. The court's denial of the preliminary injunction, the confidential settlement, and the subsequent fundraise together form a data point: in the current humanoid-dexterity race, a credible technical demonstration and investor conviction can outweigh a trade-secrets claim — even one brought by Tesla.

The Chill on Talent Mobility

The Tesla-Proception settlement did more than close a docket. It sent a signal through humanoid robotics labs in the Bay Area and beyond: the cost of switching employers just rose, even in California where non-competes are void on paper. The year-long litigation, filed June 11, 2025, dismissed with prejudice June 2, 2026, still forced a ten-person startup to burn legal capital before shipping its first production hands. For engineers watching, the message is clear: trade-secret claims can stall a career move even when non-competes cannot.

California's Business and Professions Code Section 16600 has voided non-compete clauses since the 1870s, a rule the state Supreme Court reaffirmed in Edwards v. Arthur Andersen LLP (2008). Yet nearly 36% of engineers and architects and 35% of workers in computer and math fields remain bound by such agreements, a 2025 analysis found. The Federal Trade Commission's April 2024 final rule banning non-competes nationwide was blocked by a Texas federal court in August 2024; the appeal remains pending. In practice, employers have shifted enforcement energy from non-competes to the Defend Trade Secrets Act of 2016, which lets companies sue in federal court and seek injunctions against employees accused of taking secrets to rivals. Federal trade-secret filings topped 1,100 last year, most brought by large firms against former employees.

That shift matters acutely in humanoid robotics, where the highest-value IP increasingly lives in software — control algorithms, reinforcement-learning policies, sensor-fusion pipelines, not in the tendon-driven mechanics that patents can cover. A 2026 JDSupra analysis of autonomous-systems IP strategy notes that "the most competitively significant innovations reside in software and data rather than in hardware," and that companies "frequently overlook protecting their IP around software-related innovations." The same report flags the sector's "high velocity of hiring and talent mobility" as a compounding risk: engineers move rapidly between competing firms, carrying institutional knowledge with them.

The Tesla complaint against Li alleged he downloaded confidential files on robotic-hand sensors days before resigning to found Proception. Whether those allegations held weight remains uncertain; the court denied an injunction, the parties settled, and the litigation template is now established. Employers are responding with layered defenses: invention-assignment agreements signed on day one, contributor IP audits before patent filings, trade-secret hygiene programs that mark confidential documents and restrict repository access, and departure protocols that include device return and data audits. Courts have grown more receptive to narrowly tailored non-disclosure and non-solicitation agreements while scrutinizing broad non-competes. The "inevitable disclosure" doctrine — enjoining an employee from a rival role because secret use is deemed unavoidable, remains a patchwork across states, but plaintiffs still invoke it.

For hiring managers, the practical effect is a slower, more documented onboarding. Zero G Talent's board data shows Boston Dynamics listed 27 salaried roles with a median band of $157k (range $78k–$206k) and added four positions in the past week alone; Zero G Talent's figures put the Principal Technical Program Manager for Actuators at $178k–$215k. According to Zero G Talent, Figure AI shows 61 salaried roles with a $220k median ($62k–$400k range) but zero new postings in the same window, a snapshot, not a trend, but consistent with a market where due diligence on IP exposure lengthens time-to-hire.

Engineers, meanwhile, face heightened accountability. Best-practice guides now advise: avoid downloading employer data before departure, review every signed agreement, consult counsel before joining a direct competitor, and disclose prior obligations transparently to the new employer. The settlement freed Proception to raise its $11 million seed round, but the year of litigation is a cost every subsequent spinout will factor into its runway.

The net result: talent still moves — California law sees to that, but the friction is measurable. Companies that systematize trade-secret protection and exit protocols will recruit faster; those that rely on vague confidentiality language will lose candidates to competitors with cleaner IP hygiene. The humanoid hand race won't pause, but the hiring sprint just added a legal warm-up lap.

Three Camps Racing for Dexterity

The humanoid dexterous hand market has fractured into three distinct competitive camps, each with a different bet on where the IP moat actually sits. The vertical integrators — Tesla, Unitree, Figure AI, and Zhiyuan (AGIBOT), treat the hand as a captive subsystem, co-designed with their whole-body control stack. The specialized suppliers — LinkerBot, Inspire Robots, Shadow Robot, DexRobot, Xynova, sell hands to anyone building a humanoid, chasing standardization and volume. And the crossover players — RoboSense, AAC Technologies, Zhaowei Mechanism, Jiangsu Leili, Ningbo Huaxiang, are repurposing automotive-grade actuators, sensors, and transmission lines they already mass-produce for EVs. The Tesla–Proception lawsuit has sharpened the lines between them.

Tesla's Optimus program remains the single largest demand signal in the sector. Musk's public target of 1 million-plus units by 2027 — backed by 2025 factory deployments in battery-cell sorting and automotive assembly, has forced every competitor to accelerate roadmaps and funding rounds. Unitree, the Chinese quadruped-to-humanoid crossover, prices its G1 at roughly €11,650, a fraction of Boston Dynamics' Atlas (~€120,000+) and Tesla's own estimated €25,800 bill of materials. That cost advantage stems from China's control of roughly two-thirds of the global humanoid-component supply chain, including coreless motors, six-axis force sensors, and rare-earth magnets. Unitree and Zhiyuan each shipped more than 5,000 humanoids in 2025; AGIBOT alone delivered 5,168 units, each requiring two hands. Global humanoid shipments exceeded 15,000 units in 2025 — a nearly sevenfold jump from 2024, implying more than 30,000 dexterous hands shipped last year.

Figure AI, meanwhile, has raised significant capital across multiple rounds and is iterating its Helix AI stack on a vertically integrated hand. Its hiring reflects that ambition: the company lists 61 salaried roles on the platform with a median band of $220k, and recent postings for Helix AI Engineers span pretraining, reinforcement learning, perception, and localization, all at $200k–$400k. Boston Dynamics, now under Hyundai, is hiring more quietly: 27 salaried roles with a $157k median, including a Teleoperations Research Engineer for Atlas and a Principal Technical Program Manager overseeing Actuators. Both companies are building hands in-house, betting that tight brain-to-fingertip coupling beats any off-the-shelf solution.

The specialized suppliers are scaling faster than the integrators expected. Inspire Robots delivered 10,000 standalone dexterous hands in 2025, up from 2,000 in 2024, at a commercial price of $3,000–$8,000, a five-to-tenfold drop from the Shadow Robot era's $100,000-plus research grade. LinkerBot's monthly shipments have broken 1,000 units across its L10, L20, and L30 lines (20+ DoF each), and the company targets 50,000–100,000 units in 2026, claiming over 80 percent of the high-DoF mass-production niche. Shadow Robot, the UK pioneer, has pivoted to an AI-hardware co-development model via its DEX-EE collaboration with Google DeepMind, a signal that the highest-end tactile sensing and control algorithms are now the differentiator, not the kinematics alone.

Crossover players are applying a "dimensional strike" to the cost structure. AAC Technologies reports roughly 80 percent self-sufficiency on key components, coreless motors and six-axis force sensors, IMUs, giving it integration and yield advantages that pure-play startups cannot match. RoboSense is using its lidar perception stack, dexterous hand hardware, and the VTLA-3D vision-language-action model into a full hand-eye-brain solution. SCHUNK, the German gripping giant, spun out SCHUNK Humanoid Robotics GmbH in 2026 to productize modular five-finger hands for industrial humanoids, moving from one-off prototypes to a scalable subsystem business. Festo maintains a dual track in pneumatics and soft robotics, while Robotiq's underactuated adaptive grippers continue to dominate the cobot retrofit market.

China's MIIT mandate — mass production by 2025, global leadership by 2027, has mobilized state capital at a scale no Western initiative matches. The National Venture Capital Guidance Fund and three regional funds have allocated CNY 1 trillion (€120 billion) over 20 years. Chinese firms supplied 57 percent of the domestic industrial robot market in 2024, overtaking foreign competitors for the first time. The dexterous hand segment mirrors that trajectory: Asia-Pacific leads with a 24.93 percent CAGR through 2031, driven by China's humanoid program, Japan's service robot ecosystem, and South Korea's semiconductor automation. North America held 37.3 percent of 2025 market share ($80.7 million), Europe 25.1 percent.

The lawsuit reverberates through this terrain in two ways. First, it signals to investors that Tesla treats hand IP as a crown-jewel asset, not just the tendon-driven 22-DoF architecture but the control algorithms and tactile-sensor fusion that make it reliable at production volume. That raises the bar for any startup claiming a "clean-room" design; due diligence will now demand provenance for every sensor calibration routine and tendon-routing patent. Second, it accelerates the split between the vertical integrators, who can absorb litigation risk as a cost of owning the stack, and the specialized suppliers, who must prove their IP is independently derived or licensed. Zhiyuan's recent decision to spin off its dexterous hand business into a separate subsidiary, widely read as a move toward industrial division of labor, suggests even the best-funded Chinese OEMs are hedging.

For engineers and operators, the competitive map is clarifying. If you want to work on the bleeding edge of hand-eye-brain co-design, the integrators (Tesla, Figure AI, Unitree, Zhiyuan, Boston Dynamics) are hiring for full-stack roles. If you want to push tactile sensing, tendon routing, or micro-actuator yield at volume, the specialized suppliers (LinkerBot, Inspire, Shadow Robot, PaXini, Daimon) and crossover players (AAC, RoboSense, SCHUNK) are scaling production lines. The lawsuit didn't create this split, but it made the IP boundaries visible.

What This Means for the Race

The patent data tells a story the headlines miss. Since 2006, roughly 20,000 humanoid robotics patents have been filed globally across 15,000 unique families, and 16,000 remain active, a ratio that signals sustained commercial conviction, not speculative filing. Nearly half that volume arrived in the last five years. China drove the surge, climbing from 271 filings in 2015 to 2,426 in 2025 before a dataset dip in 2026 that reflects publication lag, not innovation slowdown. The United States peaked at 170 in 2018 and fell to 91 in 2025. Japan holds steady around 2,200 total. South Korea sits at 1,100. Northeast Asia now owns the paper terrain.

The concentration is stark. UBTECH leads with 521 patents. Honda follows at 372. Toyota, Sony, SoftBank, Samsung, and Boston Dynamics each hold 150–210. Chinese startups Fourier and AgiBot have cracked the top ten with 150 and 109 respectively. Among Western startups, only Sanctuary AI reaches the global top 20. Industrial giants — Sony, UBTECH, Honda, Toyota, Hyundai, Alphabet, Samsung, collectively control more than 11,000 patent families. That wall of prior art creates a licensing gauntlet for any new entrant.

Tesla's Optimus program illustrates the deployment reality. Over 1,000 units operated across Gigafactory Texas and Fremont as of January 2026, handling parts processing, kitting, and intricate assembly. Gen 3 production starts this year. Proception, post-settlement, is shipping its first batch of 22-DoF tendon-driven hands to researchers and opening wider orders, with a roadmap targeting a data platform in 2026 and a full humanoid prototype by 2027. The market projection — $38 billion by 2035, long-term estimates beyond $7 trillion, assumes these machines actually work in unstructured environments. That assumption is the bet.

The Tesla–Proception case tested whether trade-secret law can police the boundary between general knowledge and protected IP when an engineer moves shops. The court denied Tesla's preliminary injunction, finding insufficient likelihood of success on the misappropriation merits. The parties stipulated to dismissal with prejudice in May 2026. The outcome suggests courts will demand concrete evidence of specific secret taking, not just proximity and similarity. But the litigation itself imposed months of distraction and legal cost on a ten-person startup, a tax on mobility that favors incumbents with legal war chests.

Will Rosellini, chief IP officer at PatentVest, said: "IP is now a gating factor for scale, licensing, and long-term value capture." His firm's analysis found most billion-dollar humanoid startups carry thin patent portfolios despite $2.1 billion in cumulative investment. The active-to-inactive ratio of roughly 13,000 to 1,700 shows established players maintain portfolios strategically, not for litigation alone, but for cross-licensing leverage and partnership currency.

Three dynamics bear watching. First, patent quality will separate performers from posers. Volume favors China, but the US and Japan historically file narrower, more commercially enforceable claims. Second, the AI-generated IP frontier arrives fast. Harvard scholar Andrew Hartwig projects 90 percent of global IP will be AI-generated within a decade. Current law recognizes only natural persons as inventors. The Thaler/DABUS test cases failed across jurisdictions. When a foundation model designs a tendon-routing geometry that outperforms human engineers, who owns it? Third, physical deployment raises the cost of error. A chatbot hallucination is embarrassing. A hand control error on a factory line injures someone. That liability gradient will force tighter integration of IP strategy, safety certification, and insurance, favoring companies that control the full stack.

The Tesla-Proception docket is closed, but the question it forced remains open on every factory floor where a robotic hand reaches for a part: whether the law can draw a line between the knowledge an engineer carries and the secrets a company owns. The hands are moving faster than the courts. The winners will own both the dexterity and the rights to deploy it.

Working in robotics? Zero G Talent tracks the openings: see every open Boston Dynamics role, browse robotics jobs, openings at Figure AI, and the people building the field.