2026-06-18 — views
Physical AI Patents & IP Landscape — Who Owns the Key AV and Humanoid Moats
Patent portfolios are the most durable moat in physical AI — mapping who owns AV sensor fusion, neural driving, and humanoid kinematics IP heading into 2026.
Article 52 in the Physical AI Benchmark Series — The IP Dimension
Investors and analysts tracking Tesla versus Waymo spend most of their time on operational metrics: weekly ride counts, disengagement rates, fleet size, revenue per vehicle. Those metrics describe what the systems do today. Patents describe what the companies own tomorrow. In physical AI, patents are not peripheral — they are structural. Unlike software-only AI, where broad algorithm patents are difficult to obtain and open-source erodes exclusivity within months, physical AI produces a multi-layer IP stack that spans sensor hardware, sensor fusion methods, neural network architectures, fleet-scale training pipelines, vehicle actuation systems, and humanoid kinematics. Each layer is independently patentable, and coverage in multiple layers simultaneously creates the kind of moat that takes a decade and billions of dollars to design around.
This article maps the current patent landscape: who holds the deepest coverage, what each portfolio covers, what recent filing trends signal about strategic bets, and where the competitive exposure is concentrated for each major player.
All figures marked (est.) are estimates based on public USPTO records, published research, and industry reporting. Patent counts and portfolio depth assessments have not been independently verified against complete USPTO data.
Section 1 — Why Patents Matter in Physical AI (and Not in Software AI)
The distinction between software-only AI and physical AI IP is structural, not merely a matter of degree. In software-only AI — a recommendation engine, a language model, an image classifier — the core innovations are mathematical methods and algorithms. US patent law excludes abstract mathematical concepts from patentability, making it extremely difficult to hold broad exclusivity over, say, attention mechanisms or RLHF. Open-source releases further erode any practical moat: once a technique is published in a paper and implemented in a public codebase, the IP leverage largely disappears.
Physical AI operates on different rules. The table below maps the primary patent categories in the AV and humanoid robotics stack, what each covers, and why coverage in that layer matters competitively.
| Category | What it covers | Why it matters |
|---|---|---|
| Sensor fusion | Methods for combining LIDAR, radar, and camera data into a unified scene representation | Core to any multi-sensor AV stack; depth coverage here locks competitors into design-around costs |
| Neural architecture | Network designs for end-to-end driving, occupancy networks, trajectory planning | Directly covers the training pipeline; Tesla and Waymo both hold significant positions |
| HD map creation | Automated methods for building and updating high-definition maps | Waymo’s HD map pipeline is a key operational differentiator; patents protect the automation methods |
| Fleet training | Systems for collecting labeled data from fleet vehicles at scale | Tesla’s shadow mode and fleet learning pipeline is the source of its data advantage; patents protect the collection and curation methods |
| Vehicle actuation | Steer-by-wire, brake-by-wire, fail-operational redundancy in pedal-free vehicles | Cybercab and next-generation AV platforms require these; without coverage, suppliers can copy the architecture |
| Humanoid kinematics | Joint designs, actuator control, balance and locomotion algorithms | Tesla Optimus, Figure, Boston Dynamics — the earliest filer wins the design space before volume manufacturing begins |
| Battery and motor | Cell chemistry, motor topology for AV and robotics power density | Tesla’s structural battery; CATL and BYD in China — power density at cost is the enabling constraint for both AV and humanoid |
The multi-layer nature of this stack is the key insight. A company that holds depth in only one layer — say, sensor fusion — can be designed around if competitors retain freedom to operate in the neural architecture and fleet training layers. A company that holds coverage across three or more layers forces any new entrant to solve multiple design-around problems simultaneously, multiplying the cost and time required to compete. Waymo has the deepest coverage in the sensor fusion layer. Tesla has the broadest coverage in fleet training and neural architecture. No single company dominates every layer today.
Section 2 — Waymo’s Patent Portfolio
Waymo (Alphabet/Google) has one of the largest autonomous vehicle patent portfolios in the world, accumulated since the Google Self-Driving Car Project launched in 2009. The portfolio is filed under multiple entity names — Google LLC, Waymo LLC, and predecessor entities — making comprehensive counting difficult from public records alone. Estimates from patent analytics firms place the Google/Waymo AV portfolio in the thousands of granted US patents as of 2026 (est.).
| Dimension | Detail |
|---|---|
| Portfolio depth | Thousands of granted US patents (est.); filed under Google LLC, Waymo LLC, and related entities across multiple decades |
| Sensor fusion | LIDAR-specific patents cover beam patterns, echo processing, and low-cost LIDAR manufacturing — most date to the pre-commercialization era when LIDAR was purely proprietary hardware |
| HD mapping IP | Automated map-building methods, map differential update systems, and ground-truth validation pipelines are all represented; these protect the automation layer, not the maps themselves |
| Simulation | Waymo’s Carcraft simulator — used for training and validation at scale — has associated process patents covering synthetic data generation and scenario construction |
| Remote assistance | Patents on scalable human monitoring systems for driverless vehicles; key for Phase 3 commercial deployment where one operator monitors many vehicles simultaneously |
| Enforcement history | 2017 trade-secret and patent lawsuit against Uber (Anthony Levandowski case), settled for approximately $245 million in Uber equity — Waymo has established it will enforce IP aggressively |
| Licensing posture | Limited public cross-licensing; Waymo has pursued selective licensing deals rather than broad platform licensing |
Notable Waymo patent areas visible in public USPTO records:
- US10048688B2 — System and method for processing LIDAR data (core sensor fusion method)
- Multiple patents covering occupancy grid generation from sparse LIDAR point clouds
- Patents on remote assistance systems enabling human monitoring of fully driverless vehicle fleets
- Carcraft simulation environment process patents covering synthetic scenario generation
The enforcement record is the critical signal here. The Uber settlement established that Waymo is not merely accumulating patents defensively — it demonstrated willingness to pursue litigation to conclusion and extract significant commercial value from IP. Any company developing a competing AV stack that overlaps with Waymo’s sensor fusion or HD mapping coverage must treat litigation risk as a real cost in its business model.
Section 3 — Tesla’s Patent Approach: The Open-Source Paradox
Tesla made one of the most widely discussed IP moves in the industry in 2014, when Elon Musk announced that Tesla would not initiate patent lawsuits against parties making good-faith use of its technology. This is broadly interpreted as Tesla adopting an open-patent posture. The reality is more strategically nuanced.
| Fact | Detail |
|---|---|
| 2014 pledge | Tesla pledged not to sue good-faith users — but did not release patents, transfer IP to the public domain, or dedicate the patents to a standards body |
| Patent portfolio | Thousands of granted US patents and pending applications across vehicle design, battery chemistry, software systems, and manufacturing (est.) |
| FSD-specific IP | Tesla files patents on neural network architectures, training methods, and data pipeline design — published under Tesla, Inc. (formerly Tesla Motors) in USPTO records |
| Dojo patents | Custom chip interconnect architectures, memory bandwidth optimization, and training cluster layout are represented in filings related to the Dojo training infrastructure |
| Optimus-specific | Actuator designs, foot and hand kinematics, and teleoperation interfaces for robot training data collection are emerging in recent patent filings (2024–2026, est.) |
| Strategic logic | The open pledge removes IP as a recruiting and publishing barrier — engineers can present Tesla work freely at academic conferences — while preserving actual patent rights for enforcement if needed |
Tesla’s key recent filing areas (2023–2026, est., based on public USPTO applications):
- End-to-end neural driving policies trained via imitation learning from human driver behavior
- Fleet-scale edge-case curation methods — the technical implementation of shadow mode, including how rare or high-value scenarios are identified and prioritized for labeling
- Humanoid robot joint designs using low-friction cable-driven actuators targeted at volume manufacturing cost structures
The open pledge is a recruiting instrument masquerading as an IP strategy. By removing the social friction of presenting Tesla work at academic conferences — where IP restrictions would normally prevent researchers from publishing details of proprietary methods — Tesla captures a disproportionate share of top ML talent without having to match the academic freedom of a university research lab. The actual patents remain intact. If Tesla ever chose to enforce them, the 2014 pledge provides no legal protection to competitors — it is a good-faith commitment with no contractual force, and Tesla has reserved the right to revoke it for bad-faith use.
Section 4 — Other Major Patent Holders in the Physical AI Stack
Waymo and Tesla dominate the narrative, but the physical AI IP landscape includes substantial positions held by Tier 1 automotive suppliers, semiconductor companies, and robotics startups. The table below maps the key players outside the two leaders.
| Company | Key IP focus | Notable patents or cases |
|---|---|---|
| Mobileye (Intel spinoff, IPO 2022) | Camera-based AV sensing, Responsibility Sensitive Safety (RSS) formal safety model | RSS model is widely cited in AV safety research; Mobileye has pursued a combined publication-plus-patent strategy to establish it as an industry standard |
| Qualcomm | Automotive SoC compute (Snapdragon Ride platform), V2X communication protocols | Dominant in radio and connectivity IP; C-V2X infrastructure is the connectivity layer for cooperative AV systems |
| NVIDIA | AV training infrastructure (DRIVE platform), Orin SoC architecture, DriveWorks SDK | Chip-level patents on the training accelerator; DRIVE platform APIs create a proprietary middleware layer |
| Bosch / Continental | Radar sensing, camera modules, AV middleware integration | Deep Tier 1 supplier patents on physical sensor components; any AV that uses commodity radar or cameras operates on their licensed technology |
| Toyota Research Institute | AV safety systems, formal verification methods, Guardian driver-assist architecture | Guardian patents describe a co-pilot model that could become a regulatory standard for supervised AV systems |
| Baidu Apollo | China-specific AV sensing, HD mapping, and vehicle control | Dominant in China market IP; limited US enforcement posture today, but PCT filings indicate expansion intent |
| Figure / 1X / Agility Robotics | Humanoid locomotion, manipulation, and mobile base kinematics | Early-stage portfolios; fewer granted patents but heavy filing activity as these companies approach commercialization |
Two companies in this list deserve special attention for their potential to shape the broader industry regardless of whether they win the commercial AV race:
Mobileye’s RSS strategy is an attempt to transform a proprietary safety model into a de facto regulatory requirement. If regulators in any major jurisdiction adopt RSS as a minimum safety standard for AV approval, every company developing competing systems would need to either license the RSS framework from Mobileye or demonstrate an equivalent formal safety guarantee — which is itself a patented method. This would convert Mobileye’s IP from a competitive advantage into a structural tax on the entire industry.
Qualcomm’s connectivity IP occupies a similarly structural position. Cellular V2X (C-V2X) communication between vehicles and infrastructure is increasingly viewed as a mandatory component of next-generation AV safety systems. Qualcomm holds significant patents in the underlying radio protocol layer. If C-V2X becomes a regulatory requirement, Qualcomm earns a royalty on every connected vehicle deployed globally — a position that does not require Qualcomm to win the AV race directly.
Section 5 — What Patent Filings Signal About Strategic Bets (2024–2026)
Patent applications typically publish 18 months after filing. This means that patent applications appearing in the public USPTO database in 2024–2026 reflect strategic decisions made in 2022–2024 — choices about which technical directions to protect before competitors understood their value. The table below maps the most significant filing trends and what they signal about each company’s strategic posture.
| Filing trend | What it signals |
|---|---|
| Tesla: heavy Optimus actuator filings (2024–2026, est.) | Tesla is committing to a specific cable-driven actuator architecture for Optimus and is protecting it before volume manufacturing begins; the filing volume suggests a genuine design freeze, not exploratory research |
| Tesla: FSD imitation learning methods | Protecting the core insight that powers v12 and v13 — that labeling driver actions directly, rather than labeling scene objects and hand-writing rules, produces better generalization; this is the architecture that made end-to-end FSD possible |
| Waymo: remote assistance optimization | Filing activity around scalable human oversight systems signals that Waymo is investing in the supervisory infrastructure needed to operate hundreds of driverless vehicles with a small remote operations team — the key cost lever for Phase 3 commercial scaling |
| Waymo: Gen 6 vehicle design | Purpose-built vehicle architecture filings (sensor integration, thermal management, structural design) indicate Waymo is protecting the vehicle platform itself, not just the software — a shift from pure software IP toward full-stack physical IP |
| Mobileye: RSS formal model extensions | Continued filing on RSS variants and formal verification methods is consistent with a standards-body strategy — Mobileye wants RSS to be cited in regulatory frameworks before it becomes contested |
| China (Baidu / SAIC / NIO) | Filing in China first, then PCT international extension; aggressive IP build-out focused on the Chinese market, where local AV deployment is accelerating. Limited US enforcement posture today but PCT filings indicate global expansion intent in the 2027–2030 window |
| Humanoid startups (Figure / Agility / 1X) | Filing volume is accelerating ahead of first commercial deployments; companies are racing to establish prior art in manipulation, grasping, and mobile base locomotion before the field matures into volume manufacturing |
The most strategically significant signal in recent filing data is the convergence between AV software patents and humanoid robotics patents at Tesla. Tesla’s Optimus program shares fundamental infrastructure with FSD: the neural network training pipeline, the fleet-scale data collection architecture, and the simulation environment used for policy development. Patents filed for Optimus actuator control that cite Tesla’s FSD training methods are creating cross-domain IP linkages that would be extremely difficult for a pure-play humanoid robotics company to replicate without also replicating the AV training stack. This structural coupling between the two programs is the clearest expression in the patent record of Tesla’s thesis that general physical intelligence — rather than specialized AV or specialized humanoid — is the actual product being developed.
What to Watch: The IP Signals That Precede Market Moves
Patent filing patterns are a leading indicator with an 18-month lag. The following are the specific signals that should accompany any future analysis of the physical AI competitive landscape.
Citation density — When a competitor’s patent application cites another company’s granted patent as prior art, it provides direct evidence of technical overlap. Rising citation density between Waymo and Tesla patents, or between Western AV companies and Chinese applicants, is a signal that design-around pressure is building.
PCT international filing rate — A company filing at high PCT rates is signaling intent to enforce globally, not just in the US. Baidu’s rising PCT filing rate is the clearest indicator that China-origin AV IP is being positioned for global assertion.
Continuation filing volume — A company that files large numbers of continuation applications (extensions of existing patents) is building claim coverage depth — trying to own not just the original invention but every commercially relevant variant of it. High continuation volume is the signature of a company preparing for litigation rather than merely protecting its own freedom to operate.
Cross-domain citation patterns — Citations linking AV neural architecture patents to humanoid kinematics patents in the same assignee’s portfolio are the clearest visible signal of a unified physical AI IP strategy. This is the pattern to watch in Tesla’s filing record over the next 12–18 months.
The physical AI patent race is not a secondary story to the operational metrics — it is the structural layer that determines which of today’s operational leaders can sustain their advantage when the competitive intensity increases. A company that is operationally ahead but IP-thin is acquirable, licensable, or litigatable. A company with deep cross-layer IP coverage has built a moat that compound over time, independent of any individual product cycle.
Sources: USPTO Patent Full-Text and Image Database (patents.google.com); Tesla open-source patent pledge, Tesla Blog 2014 (tesla.com/blog/all-our-patent-are-belong-you); Waymo vs. Uber trade secret lawsuit, Reuters; Mobileye Responsibility Sensitive Safety model (mobileye.com/technology/rss/). All figures marked (est.) are estimates based on public patent records, published research, and industry reporting; they have not been independently verified against complete USPTO data and should be treated as directional rather than precise.
Sources
- USPTO Patent Full-Text and Image Database ↗
- Tesla open-source patent pledge — Tesla blog 2014 ↗
- Waymo vs Uber trade secret lawsuit — Reuters ↗
- Mobileye RSS safety model — Mobileye ↗