In a decision that underscores the importance of prior art in the context of AI patents, the PTAB recently issued a final decision in Tesla, Inc. v. Autonomous Devices, LLC, IPR2023-01173 (PTAB January 3, 2025), invalidating all challenged claims of U.S. Patent Number 11,055,583 (the “’583 patent”). The case provides some insight into how the PTAB evaluates obviousness in AI patents.
Broadly speaking, the ’583 patent relates to “artificially intelligent devices, systems, and methods for learning and/or using visual surrounding for autonomous device operation.” The ’583 patent, column 59, lines 18-20. This enables “learning one or more digital pictures of a device’s surrounding along with correlated instruction sets for operating the device, storing this knowledge in a knowledgebase (i.e. neural network, graph, sequences, etc.), and autonomously operating a device.” Id. at column 59, lines 14-18.
Claim 10 of the ’583 patent recites a system with two learning (e.g., training) processes, with a device operated with different users during each process. Claim 10 further includes that the training data for both processes is included in a knowledgebase. Petitioner argued that these features would have been obvious over U.S. Patent No. 9,604,359 (“Grotmol”), which relates to “training and/or operating a robotic device to follow a trajectory.” Grotmol, Abstract.
The Patent Owner argued that Grotmol taught only one learning process performed by one user using one device. The PTAB did not agree. Grotmol taught that its “system may be configured to learn from user commands to control the environment (e.g. robot).” Grotmol, column 25, lines 30-31. The PTAB equated this to a first learning process. Grotmol further taught “After is user is done training a particular behavior, he may create another module and activate it…The system may be configured to learn the second behavior from the user.” Id. at column 26, lines 1-6. The PTAB equated this activation of another module to the claimed second learning process, thus finding that Grotmol taught two learning processes.
The PTAB also found that having two different users, one for each learning process, using a single training device would have been rendered obvious by Grotmol. Specifically, the PTAB reasoned that Grotmol taught that the robot could be taught “simple” and “complex” tasks. Grotmol, column 24, lines 17-23. And the PTAB determined that “based on Grotmol’s teaching that a robot can learn both simple and complex tasks, a person of ordinary skill in the art would have had a specific reason for different users to train Grotmol’s robot for different behaviors, namely, to benefit from the users’ varying skills and knowledge.” Tesla at 12. Even further, the PTAB reasoned that Grotmol taught training the robot “using a gamepad,” and that switching the gamepad between users would have been simple. Tesla at page 12.
Claim 11 is different than claim 10 because claim 11 uses different devices during the first and second learning processes. But, the PTAB found that this was also obvious over Grotmol because expert testimony indicated that a person having ordinary skill (“POSA”) would have known to load a robot with additional configurations trained on other robots to maximize performance. Id. at 16.
Claim 25 includes a server that facilitates the transfer of training data (e.g., in this case digital pictures) or instruction sets between devices. The PTAB found this feature obvious over Grotmol in view of U.S. Patent No. 8,639,644 (“Hickman”). More specifically, Grotmol taught loading images correlated with commands from a first robot to a second robot. Tesla at page 18. Hickman was then cited for its disclosure of a shared robot knowledge base. The PTAB concluded that, in the combined Grotmol-Hickman system, the training data would have been sent from the first robot to the knowledge base, and then to the second robot.
The PTAB also concluded that there would have been reason to combine Grotmol and Hickman, stating: “…it would have been obvious to a person of ordinary skill in the art to combine Grotmol and Hickman so that Grotmol’s robots share training sets using Hickman’s shared robot knowledgebase…Doing so would have been obvious because Grotmol teaches sharing training sets between robots, and Hickman’s shared robot knowledgebase would facilitate wider and faster sharing of Grotmol’s training sets.” Id. at 24.
The Patent Owner argued that Hickman taught away from the proposed combination. Specifically, the Patent Owner argued that Hickman taught instruction sets and images (e.g., training data) stored in separate knowledge bases, whereas Grotmol’s training set included images and instruction sets that were correlated and stored together. The Patent Owner thus argued that the two systems were incompatible. But, the PTAB responded: “Even if Patent Owner is correct that Hickman’s Figure 3 shows images stored separately from instruction sets, Hickman’s Figure 3 is just an ‘example shared robot knowledge base.’” Id. at 25, quoting Hickman, column 11, line 42. The PTAB additionally reasoned: “Hickman teaches that its shared robot knowledgebase shown can store an image correlated with an instruction set. For example, Hickman teaches that general object knowledge base 308 in Figure 3 includes ‘image data (IMG)’ correlated with ‘task data (TASKS).’”
Furthermore, the PTAB explained that there was motivation to combine, stating, “a person of ordinary skill in the art would have combined Grotmol and Hickman because Hickman’s shared robot knowledgebase would facilitate wider and faster sharing of Grotmol’s training sets.” Id. at page 26.
Claim 30 included generating a picture including a top-down view of an object. According to Petitioner’s expert, this was taught by U.S. Patent No. 9,283,674 (“Hoffman”). The Patent Owner argued that there was no motivation to combine Grotmol and Hoffman because Grotmol taught autonomous operation, whereas Hoffman taught manual operation. The PTAB disagreed, stating that both Grotmol and Hoffman disclosed both semi-autonomous and autonomous robots. The PTAB further stated that a POSA “would have had a specific reason to combine Grotmol and Hoffman because Hoffman’s top-down view would benefit a user by providing more detailed information about the surroundings of Grotmol’s robot.” Id. at 34.
Tesla may show that in AI patents dependent claims are vulnerable when they add only small, incremental limitations. For instance, claims specifying multiple users or devices were deemed to be obvious over single-user or single-device systems. For example, claim 10, which specified two users (one for each learning process), was deemed obvious over Grotmol. In another example, regarding the AI training, claim 10 recited two learning processes. However, the PTAB broadly construed Grotmol to conclude that Grotmol also taught two AI training processes. In drafting claims, practitioners should consider whether dependent claims add novel, nonobvious limitations, or merely recite predictable implementations.