DCT

2:25-cv-00742

Perceptive Automata LLC v. Tesla Inc

Log In

I. Executive Summary and Procedural Information

  • Parties & Counsel:
  • Case Identification: 2:25-cv-00742, E.D. Tex., 07/23/2025
  • Venue Allegations: Plaintiff alleges venue is proper because Defendant Tesla has regular and established places of business within the Eastern District of Texas and has committed acts of infringement in the district, including the making, using, and selling of vehicles with the accused technology.
  • Core Dispute: Plaintiff alleges that Defendant’s autonomous vehicle systems, including its Full Self-Driving (FSD) technology, infringe five patents related to using machine learning models to predict the behavior and intent of humans in traffic environments.
  • Technical Context: The technology at issue involves advanced driver-assistance systems that use sensor data and artificial intelligence to anticipate the actions of pedestrians, cyclists, and other drivers, aiming to improve autonomous vehicle safety and performance.
  • Key Procedural History: The complaint alleges that Tesla was aware of the asserted patents prior to the lawsuit, citing numerous instances where Tesla referenced the patents during the prosecution of its own patent applications, a factor that may be relevant to the allegations of willful infringement.

Case Timeline

Date Event
2017-07-05 Earliest Priority Date for ’344, ’889, and ’046 Patents
2019-01-30 Earliest Priority Date for ’346 Patent
2019-02-06 Earliest Priority Date for ’579 Patent
2020-04-07 U.S. Patent No. 10,614,344 Issues
2021-09-21 U.S. Patent No. 11,126,889 Issues
2022-10-11 U.S. Patent No. 11,467,579 Issues
2022-12-06 U.S. Patent No. 11,520,346 Issues
2023-09-12 U.S. Patent No. 11,753,046 Issues
2024-01-01 Tesla re-incorporates from Delaware to Texas
2025-07-23 Complaint Filed

II. Technology and Patent(s)-in-Suit Analysis

U.S. Patent No. 10,614,344 - "System and Method of Predicting Human Interaction with Vehicles" (Issued Apr. 7, 2020)

The Invention Explained

  • Problem Addressed: The patent describes that conventional autonomous driving systems are deficient because they predict a person's future movements based only on their past movements (i.e., motion vectors), which is insufficient for anticipating complex human behavior in urban environments (’344 Patent, col. 1:42-50).
  • The Patented Solution: The invention proposes a system where an autonomous vehicle's computing device receives sensor data (e.g., video) of an object on the road. This data is fed into a supervised learning model, which is configured to predict a "statistical summary characterizing a distribution of user responses" associated with that object. The vehicle's operation is then controlled based on this generated statistical summary data (’344 Patent, Abstract; col. 1:52-2:2). This approach aims to model how a collection of human observers would interpret and react to the scene, rather than relying solely on physics-based motion prediction.
  • Technical Importance: The technology represents a shift from purely kinematic predictions to a psychometric approach, attempting to computationally model human intuition about the "state of mind" of other road users to enable safer and more natural vehicle operation (’344 Patent, col. 7:32-39).

Key Claims at a Glance

  • The complaint asserts at least independent claim 1 (Compl. ¶23).
  • Essential elements of Claim 1 include:
    • Receiving sensor data from an autonomous vehicle displaying an object on the road.
    • Providing the sensor data as input to a supervised learning based model, which is configured to predict an output statistical summary characterizing a distribution of user responses.
    • Executing the trained model to generate the statistical summary data.
    • Controlling the operation of the autonomous vehicle based on the generated statistical summary data.
  • The complaint reserves the right to assert additional claims (Compl. ¶23 n.1).

U.S. Patent No. 11,126,889 - "Machine Learning Based Prediction of Human Interactions with Autonomous Vehicles" (Issued Sep. 21, 2021)

The Invention Explained

  • Problem Addressed: Similar to its parent, this patent addresses the inability of autonomous systems to adequately predict human behavior, noting that computers lack the natural human ability to effortlessly predict a person's actions (’889 Patent, col. 1:39-44).
  • The Patented Solution: This invention focuses on the method of creating the predictive model. The method involves receiving images of road scenes and a plurality of "user responses, each user response describing a state of mind of a road user" displayed in the images. A training dataset is generated from these responses, which is then used to train a supervised learning model. An autonomous vehicle then uses this trained model to receive a new image, predict summary statistics describing the road user's state of mind, and control the vehicle based on that prediction (’889 Patent, col. 15:62-16:21).
  • Technical Importance: This patent details the specific process of training a model on human-labeled "state of mind" data, providing a technical framework for converting qualitative human judgments into a functional control system for an autonomous vehicle (’889 Patent, col. 7:38-44).

Key Claims at a Glance

  • The complaint asserts at least independent claim 2 (Compl. ¶39).
  • Essential elements of Claim 2 include:
    • Receiving a plurality of images of road scenes.
    • Receiving a plurality of user responses, each describing a state of mind of a road user in an image.
    • Generating a training dataset from the user responses.
    • Training a supervised learning based model with the dataset.
    • In an autonomous vehicle, receiving a new image of a scene including a road user.
    • Predicting, with the model, summary statistics describing the state of mind of the road user.
    • Controlling the vehicle based on the prediction.
  • The complaint reserves the right to assert additional claims (Compl. ¶23 n.1).

U.S. Patent No. 11,467,579 - "Probabilistic Neural Network for Predicting Hidden Context of Traffic Entities for Autonomous Vehicles" (Issued Oct. 11, 2022)

  • Technology Synopsis: This patent describes using a probabilistic neural network to predict "hidden context" (e.g., intentions, future plans) for traffic entities. The system generates output values representing this context and also determines a measure of uncertainty for those values, which is then used to navigate the vehicle, for example, by adjusting its path planning (’579 Patent, Abstract; col. 2:35-40).
  • Asserted Claims: At least independent claim 1 (Compl. ¶59).
  • Accused Features: The complaint alleges Tesla's FSD software is a probabilistic neural network that generates output representing hidden context (e.g., predicting an oncoming car will yield or go around parked cars) and uses this to navigate the vehicle (Compl. ¶62, ¶65, ¶69).

U.S. Patent No. 11,520,346 - "Navigating Autonomous Vehicles Based on Modulation of a World Model Representing Traffic Entities" (Issued Dec. 6, 2022)

  • Technology Synopsis: This patent claims a method where an autonomous vehicle generates a point cloud representation of its surroundings, identifies traffic entities, and uses a machine learning model to determine a "hidden context" for each entity. Based on this context, the system modifies a region of the point cloud where the entity is expected to travel and navigates the vehicle to stay a threshold distance away from this modified region (’346 Patent, col. 1:45-2:4).
  • Asserted Claims: At least independent claim 1 (Compl. ¶80).
  • Accused Features: The complaint alleges Tesla's FSD generates a 3D point cloud, identifies traffic entities, determines their hidden context (e.g., predicted path), modifies a region based on that context, and navigates to maintain a safe distance from that modified region (Compl. ¶83, ¶87-90).

U.S. Patent No. 11,753,046 - "System and Method of Predicting Human Interaction with Vehicles" (Issued Sep. 12, 2023)

  • Technology Synopsis: This patent is directed to a method of generating training data for a predictive model. The process involves storing images, sending them to human observers with a question about a user's "state of mind" in the image, receiving responses, and generating summary statistics from those responses. This training data is then used to train a model that predicts a user's state of mind in a new image (’046 Patent, Abstract; col. 15:44-16:2).
  • Asserted Claims: At least independent claim 1 (Compl. ¶101).
  • Accused Features: The complaint alleges Tesla's "Data Engine" stores images from its vehicle fleet and uses a team of over 1,000 data labelers who provide answers about the state of mind of road users, thereby generating training data with summary statistics to train the FSD model (Compl. ¶103-108).

III. The Accused Instrumentality

Product Identification

The accused instrumentalities are Tesla vehicles equipped with Full Self-Driving (FSD) hardware and software, specifically including the Model Y and Cybertruck models (Compl. ¶15). The complaint also implicates the backend systems used to train the FSD models, such as the "Cortex" and "Dojo" supercluster datacenters (Compl. ¶16, ¶24, ¶40).

Functionality and Market Context

The complaint alleges that Tesla's FSD system uses an onboard computer and sensors, such as cameras, to monitor road conditions and run a "Tesla-developed neural net" (Compl. ¶25). This system is designed to perform autonomous driving functions by processing vast amounts of sensor data to perceive the world and predict the actions of other road users. The complaint cites Tesla's public statements that its neural network "learns from millions of examples of what humans have done," characterizing the system as "human imitation" (Compl. ¶27). The complaint also points to Tesla's large-scale data collection from its fleet of vehicles, which is used to continuously train and improve the FSD models through its "Data Engine" (Compl. ¶42, ¶65). A diagram from a Tesla AI Day presentation illustrates this "Data Engine" pipeline, showing data flowing from the "Tesla Fleet" through training sets to deploy updated models (Compl. ¶42, p. 21).

IV. Analysis of Infringement Allegations

'344 Patent Infringement Allegations

Claim Element (from Independent Claim 1) Alleged Infringing Functionality Complaint Citation Patent Citation
receiving, by a computing device associated with an autonomous vehicle... sensor data captured by a sensor installed on the autonomous vehicle, the sensor data displaying an object on the road; Tesla vehicles have an onboard computer that receives sensor data from cameras that monitor road conditions and display objects like other vehicles and pedestrians. ¶25 col. 15:26-31
providing the sensor data as input to a supervised learning based model, the supervised learning based model configured to... predict... an output statistical summary characterizing a distribution of user responses expected to be received... Tesla's FSD software is a supervised learning model. It is presented with sensor data and creates "goal candidates" that "correspond to a probability mask derived from human demonstration," which is alleged to be a distribution of user responses. ¶26 col. 15:32-38
executing the trained supervised learning based model to generate a statistical summary data characterizing a distribution of user responses... The FSD software executes the model to generate its predictions, allegedly evidenced by its assessment of "Intervention Likelihood" and its use of a "Human-like Discriminator," which are based on "millions of examples of what humans have done." ¶27 col. 15:39-44
controlling the operation of the autonomous vehicle on the road based on the generated statistical summary data. The FSD software uses the generated statistical data to determine actions, such as whether to "Assert to Pedestrian" or "Yield to Pedestrian." An "Interaction Search" diagram from Tesla shows this decision-making process. ¶28; p. 14 col. 15:45-47
  • Identified Points of Contention:
    • Scope Questions: A central question may be whether Tesla’s system, which is described as learning through "human imitation" (Compl. ¶27), generates a "statistical summary characterizing a distribution of user responses" as required by the claim. The dispute may focus on whether the output of Tesla's neural network, such as "goal candidates" or an "Intervention Likelihood," meets the specific definition of a "distribution of user responses" as contemplated by the patent.
    • Technical Questions: The complaint alleges that a "probability mask derived from human demonstration" satisfies the "distribution of user responses" limitation (Compl. ¶26). A key technical question will be what evidence demonstrates that this "probability mask" is functionally and structurally equivalent to the claimed "distribution of user responses," which the patent describes as an estimate of what a "collection of human observers" would say (’344 Patent, col. 13:9-15).

'889 Patent Infringement Allegations

Claim Element (from Independent Claim 2) Alleged Infringing Functionality Complaint Citation Patent Citation
receiving a plurality of images displaying road scenes captured by one or more vehicles; Tesla's datacenters (Dojo/Cortex) receive a plurality of images and video clips displaying road scenes from its fleet of customer vehicles. ¶42 col. 15:43-45
receiving a plurality of user responses, each user response describing a state of mind of a road user displayed in one or more images; Tesla employs over 1,000 human labelers who assess video clips and provide labels (e.g., selecting from options like "stopped_traffic") that allegedly describe the "state of mind" of the road user. ¶43 col. 15:46-49
generating a training dataset comprising summary statistics of user responses describing the state of minds of road users... Tesla generates a training dataset at its datacenters that comprises summary statistics from the labelers' responses, which is used to improve "Vehicle Movement Accuracy" over time. A diagram shows this "Data Engine" process. ¶44; p. 23 col. 15:50-52
training, using the training dataset, a supervised learning based model configured to predict summary statistics describing a state of mind of a road user... Tesla uses the generated training dataset to train its FSD software, which is a supervised learning model that allegedly predicts actions based on the "possible state of mind of road users." ¶45 col. 15:53-56
receiving, by an autonomous vehicle, a new image captured by a camera... the new image of a scene including a road user; A Tesla vehicle, during operation, receives a new image from its camera showing a road user, such as a pedestrian. ¶46 col. 15:57-59
predicting, by the autonomous vehicle... summary statistics describing a state of mind of the road user in the new image; The FSD software makes statistical predictions describing the state of mind, evidenced by its ability to predict whether a road user will turn, continue forward, or cross the road. ¶47 col. 15:57-59
controlling the autonomous vehicle based on the prediction of the supervised learning based model. The vehicle's operation is controlled based on the prediction, such as stopping to allow a pedestrian to cross the road. ¶48 col. 15:60-61
  • Identified Points of Contention:
    • Scope Questions: The case may turn on the definition of "state of mind." Tesla's system is alleged to receive labels describing a vehicle's situation (e.g., "stopped_traffic") (Compl. ¶43). The question is whether providing such situational labels constitutes describing a "state of mind" (e.g., intent, awareness) as required by the claim language.
    • Technical Questions: What evidence does the complaint provide that the labels generated by Tesla's human reviewers are used to create "summary statistics" that are then directly predicted by the in-vehicle model? The complaint shows a graph of improving accuracy over time (Compl. ¶45; p. 24), but the connection between this aggregate improvement and the specific claim requirement of predicting "summary statistics" may be a point of dispute.

V. Key Claim Terms for Construction

For the ’344 Patent

  • The Term: "statistical summary characterizing a distribution of user responses"
  • Context and Importance: This term is the core of the invention's departure from simple motion prediction. Its definition will determine whether Tesla's "human imitation" model, which learns from vast amounts of driving data, falls within the claim scope. Practitioners may focus on this term because the infringement theory hinges on equating Tesla's "probability mask derived from human demonstration" (Compl. ¶26) with this specific claim language.
  • Intrinsic Evidence for Interpretation:
    • Evidence for a Broader Interpretation: The specification states the goal is to predict behavior in a way that "more closely resembles how human drivers would predict their behavior" (’344 Patent, col. 3:44-45), which could support interpreting the term to cover any model output that aggregates human-like judgments, not just formal statistical measures.
    • Evidence for a Narrower Interpretation: The specification details that the "summary statistics are an estimate of what the summary statistics would be for a collection of human observers" shown the scene (’344 Patent, col. 13:9-13). This could support a narrower construction requiring an explicit simulation or prediction of how a group of people would vote or respond, potentially including formal statistical parameters like variance or skew (’344 Patent, col. 6:15-20).

For the ’889 Patent

  • The Term: "state of mind"
  • Context and Importance: This term is central to the training method claimed. The infringement allegation rests on the idea that Tesla's human data labelers are describing a "state of mind," not just a physical situation. The construction of this term will be critical to determining whether Tesla's data-gathering and labeling process infringes.
  • Intrinsic Evidence for Interpretation:
    • Evidence for a Broader Interpretation: The patent states that the system improves performance by allowing vehicles "to make judgments about the future behavior of road users based on their state of mind" (’889 Patent, col. 7:41-44). This suggests "state of mind" could be broadly construed to encompass any internal cognitive state that leads to a future behavior, which might include the situational assessments made by Tesla's labelers.
    • Evidence for a Narrower Interpretation: The specification gives specific examples of what judgments a human observer might make, including assessments of "intention, awareness, personality, state of consciousness, level of tiredness, aggressiveness," etc. (’889 Patent, col. 9:35-40). This language could support a narrower construction requiring the user responses to describe a subjective mental or emotional state, rather than an objective description of a vehicle's action (e.g., "stopped_traffic").

VI. Other Allegations

  • Indirect Infringement: The complaint alleges inducement of infringement under 35 U.S.C. § 271(b) for all asserted patents. The basis for this allegation is that Tesla advertises and provides instructions, user manuals, and marketing for its FSD features, which allegedly "facilitate, direct or encourage the use of infringing functionality with knowledge thereof" (Compl. ¶12, ¶33, ¶53).
  • Willful Infringement: The complaint alleges that Tesla's infringement has been and continues to be willful. This is primarily based on allegations of pre-suit knowledge, asserting that Tesla knew of the asserted patents because they were cited during the prosecution of numerous Tesla patents before the USPTO (Compl. ¶32, ¶52, ¶113). The complaint alleges that despite this knowledge, Tesla continued its infringing conduct, disregarding an objectively high likelihood of infringement (Compl. ¶34, ¶54).

VII. Analyst’s Conclusion: Key Questions for the Case

  • A core issue will be one of definitional scope: can terms like "state of mind" and "distribution of user responses," which imply modeling subjective human cognition, be construed to cover a neural network trained on vast quantities of objective driving data described as "human imitation"? The outcome may depend on whether the court views Tesla's system as merely mimicking behavior versus performing the specific psychometric analysis described in the patents.
  • A key evidentiary question will be one of technical implementation: what evidence will demonstrate that the internal workings of Tesla's FSD software and "Data Engine" perform the specific steps recited in the claims? While the complaint cites Tesla's high-level presentations, the case will likely require a deeper technical showing that Tesla's process of data labeling, model training, and in-vehicle prediction maps directly onto the patented methods.