US 10402687

I'll research this patent number across the sources you mentioned.

Let me dig into the claims text and the CAFC 2026 docket activity.

The district court and PTAB activity is clear. Let me now pin down the exact claim text for the '687 patent and check for any Federal Circuit 2026 appeal.

US 10,402,687 — Analyst Summary

Caution flag up front: I searched for the literal number 10402687 / US10402687B2. I found no Federal Circuit (CAFC) 2026 docket naming this patent. What I did find is district-court and PTAB activity involving related Perceptive Automata patents (not the '687 itself). Details and confidence levels below.

---

1. Bibliographic data (authoritative: Google Patents page you supplied + Justia)

Data discrepancy noted: A third-party aggregator (Unified Patents portal) lists the priority date as 2017-07-04 and the application date as 2017-12-03. That conflicts with the patent's own face and the USPTO filing receipt (Dec. 4, 2017 filing; Jul. 5, 2017 provisional). I treat the aggregator entry as erroneous and the patent/Google Patents data as authoritative.

Unverified: I saw a government-interest statement ("NSF Award No. 1738479, Division of Industrial Innovation and Partnerships") in the later family member US 11,753,046. I could not confirm from the text I retrieved that the identical statement appears in the '687, though it is plausible given the common specification.

---

2. Abstract (verbatim, as published)

Systems and methods for predicting user interaction with vehicles. A computing device receives an image and a video segment of a road scene, the first at least one of an image and a video segment being taken from a perspective of a participant in the road scene and then generates stimulus data based on the image and the video segment. Stimulus data is transmitted to a user interface and response data is received, which includes at least one of an action and a likelihood of the action corresponding to another participant in the road scene. The computing device aggregates a subset of the plurality of response data to form statistical data and a model is created based on the statistical data. The model is applied to another image or video segment and a prediction of user behavior in the another image or video segment is generated.

---

3. Plain-language overview of the independent claims

I retrieved the full text of claim 1 and the preamble/body of claim 11. I could not independently retrieve the tail of claim 11 or any third independent claim (likely a computer-readable-medium claim near the end of the 19-claim set), so treat the third-claim description below as inferred, not verified.

Claim 1 — "A computerized method of predicting user interaction with vehicles…" (method claim)

The preamble frames the purpose: letting motorists, cyclists and pedestrians anticipate each other's actions. The steps, in order:

1. Receive an image and/or video segment of a road scene, captured from the perspective of a participant in the scene (i.e., the ego-vehicle/driver viewpoint), where the scene includes a pedestrian, cyclist, or motor vehicle.
2. Generate "stimulus data" from that image/video. The stimulus data is either the image/video itself or an altered version of it.
3. Transmit the stimulus data to a user interface — i.e., show it to human observers (e.g., crowdworkers).
4. Receive response data back from the user interface, containing human-inputted judgments about an action, or the likelihood of an action by the pedestrian/cyclist/vehicle.
5. Aggregate a subset of those responses for a given image/video into statistical data (the aggregate human judgment).
6. Create a model based on that statistical data (i.e., train a model to reproduce the aggregate human judgment).
7. Apply the model to a second image/video segment (typically live/real-world footage).
8. Generate a prediction of user behavior in that second image/video segment.

In short: human crowdsourcing on modified road-scene clips → aggregate the crowd's judgments into a numeric target → train a model on it → run the model on live driving footage to predict what the road user will do. The key differentiator asserted over prior art is that the model is trained on human predictors' judgments rather than only on extrapolated motion vectors.

Claim 11 — "A computing system… comprising a memory containing instructions for execution by a processor…" (system claim)

The same eight-step functional sequence, but recited as processor-configured instructions: generate stimulus data (original or altered image/video) → transmit to a user interface → receive the plurality of response data (action / likelihood of action for the pedestrian, cyclist or motor vehicle) → aggregate a subset into statistical data → create a model → apply the model to a second image/video → generate a prediction of user behavior. Essentially claim 1 recast in system/apparatus form.

Likely third independent claim (~claim 15) — UNVERIFIED

Given the 19-claim structure and the pattern used across the family (independent method + system + non-transitory computer-readable-medium claims), I expect a third independent claim in computer-readable-medium form mirroring claim 1. I did not retrieve its text; do not rely on this without checking the printed claim set.

Notable dependents (context for the independent claims)

• Cl. 2 — the model is a supervised learner: random forest regressor, support vector regressor, neural network, deep CNN, recurrent NN, or LSTM.
• Cl. 3–4 — generating stimulus data by manipulating pixel/array data and creating a manipulated file; specifically a looped version of extracted frames.
• Cl. 5 — the action is staying in place, changing lanes, or crossing a street; the likelihood is an ordinal value.
• Cl. 6–7 — statistical data parameters: content of response, response time, eye position relative to the display; characterized by central tendency, variance, skew, kurtosis, scale, histogram.
• Cl. 8–9 — the second image/video is from a live road scene, received from a vehicle participating in it.
• Cl. 10 — if the second segment is the same as the first, compare prediction vs. statistical data to produce an error signal and adjust model weights (i.e., a feedback/retraining loop).

---

4. Litigation and 2026 docket status

USPTO/PTAB: The patent family has drawn a large IPR campaign by Tesla, Inc. (IPR2025-01573 through IPR2025-01577, all filed Sept. 30–Oct. 10, 2025; Tesla, Inc. v. Perceptive Automata LLC). At least one, IPR2025-01573, was reported as instituted on 03/17/2026, with institution decisions for the others expected around April 7 and April 16, 2026. Perceptive filed Patent Owner Preliminary Responses and discretionary-denial briefs invoking Fintiv and "settled expectations." Tesla's asserted grounds include WO2014/210334 (Cox et al.), a reference naming Perceptive inventor Samuel Anthony.

District court: Perceptive Automata LLC v. Tesla, Inc., No. 7:25-cv-00594 (W.D. Tex., Midland-Odessa Div., Judge Derek T. Gilliland). Docket highlights: Tesla opposed motion to stay (Oct. 14, 2025); Perceptive filed a First Amended Complaint (Mar. 5, 2026); Tesla moved to dismiss under Rule 12(b)(6)/(8)(a)(2) (Apr. 2, 2026); a joint motion to stay was granted Apr. 9, 2026, and a docket entry dated June 25, 2026 shows the civil case terminated. An earlier E.D. Tex. case (2:25-cv-00742, filed Dec. 29, 2025) was voluntarily dismissed by Perceptive.

Important caveat: the patents actually asserted in that campaign appear to be US 10,614,344; 11,126,889; 11,467,579; 11,520,346; and 11,753,046 — i.e., descendants of the '687. I did not find the '687 itself named as an asserted patent in the Tesla case, and I found no CAFC appeal docket in 2026 involving US 10,402,687. If you need a definitive CAFC negative, that should be confirmed against PACER's CAFC docket search rather than inferred from web results.

---

5. Explicit uncertainty / confidence

• High confidence: title, inventors, assignee chain, application number, filing date, issue date, priority provisional, abstract, and the text of claim 1 and the opening of claim 11 (cross-checked against Justia's claim listing, which matches the specification excerpt).
• Moderate confidence: the exact scope/limits of the third independent claim; the presence of the NSF government-interest statement in the '687 specifically.
• Not established: any Federal Circuit 2026 docket for this patent number. The closest verified CAFC-adjacent activity is the PTAB IPR campaign and the now-stayed/terminated W.D. Tex. action.
• The patent's legal status is shown as Active in Google Patents, but a security interest in favor of Piccadilly Patent Funding LLC (2025) is of record — status assertions are not legal conclusions.

Sources: patents.google.com/patent/US10402687B2/en (specification, abstract, classification, assignee history); patents.justia.com/patent/10402687 (claim 1 and dependents); ptacts.uspto.gov PTAB petition documents (claim charts, IPR party positions, docket excerpts); courtlistener.com docket 72086670 (Perceptive Automata LLC v. Tesla, Inc., 7:25-cv-00594); uspto.report/patent/grant/10402687; freepatentsonline.com/10402687.html.

I'll research litigation involving US patent 10402687. Let me search multiple sources.

Let me dig deeper for any cases specifically asserting the '687 patent, and check aggregators.

Bottom line

I found no litigation in which U.S. Patent No. 10,402,687 (the '687 patent) itself has been asserted as an asserted patent. The '687 patent is a family member — specifically the parent/priority application — of the patents that Perceptive Automata is actually litigating, but the '687 patent does not appear on any infringement contentions, complaint exhibit list, or IPR challenge that I could locate.

What exists is family-level litigation involving Perceptive Automata's continuation patents, not the '687 patent. Here is what I found, with that important distinction flagged throughout.

---

1. The '687 patent's role in the litigation (why it appears in searches but is not asserted)

The '687 patent (App. No. 15/830,549, filed Dec. 4, 2017, priority to provisional 62/528,771 filed July 5, 2017, granted Sept. 3, 2019, assignee Perceptive Automata LLC) is the earliest-filed non-provisional in the family.

Per Perceptive's own PTAB briefing, the patents that are being asserted in litigation claim priority to the '687 patent. Quoting the Patent Owner's Brief in IPR2025-01574:

"The challenged patent (U.S. Patent No. 11,126,889 …) claims priority to U.S. Patent Application No. 16/512,560, filed on July 16, 2019 (and which issued as U.S. Patent No. 10,614,344 …) and U.S. Patent Application No. 15/830,549, filed on December 4, 2017 (and which issued as U.S. Patent No. 10,402,687), as well as to U.S. Provisional Patent Application No. 62/528,771 …"

So the '687 patent is cited in the litigation record as a priority document / family parent, not as an asserted patent. This is why a keyword search for "10,402,687" surfaces litigation documents even though the patent is not being enforced.

The Google Patents family-listing "Family has litigation" flag and the Darts-IP link (family=64903265) refer to this family-level litigation, not to an assertion of the '687 patent specifically.

---

2. The actual litigation: Perceptive Automata LLC v. Tesla, Inc.

This is the only litigation I could identify involving this patent family. It was filed twice (venue-driven refiling), and the asserted patents in both versions were:

U.S. Patent Nos. 10,614,344; 11,126,889; 11,467,579; 11,520,346; and 11,753,046

— none of which is the '687 patent.

(a) Original suit — Eastern District of Texas

Docket control order (Oct. 10, 2025) had set Markman for Nov. 12, 2026 and jury selection for May 17, 2027 before the case was refiled elsewhere.

(b) Refiled suit — Western District of Texas

---

3. Related PTAB proceedings (not district-court litigation, but part of the same dispute)

Tesla filed five IPRs against the five asserted patents:

Again, the '687 patent is not among the patents challenged in any of these IPRs.

---

4. Important caveats and limitations

1. No assertion of the '687 patent found. Based on searches of the Google Patents family record, PTAB filings, CourtListener/RECAP dockets, Justia dockets, PlainSite, PACER Monitor, Docket Alarm, and Unified Patents-style coverage, I could not identify any complaint, IPR, or other proceeding in which U.S. 10,402,687 is an asserted or challenged patent. If the '687 patent has been asserted, it was not surfaced by the sources I could reach.

2. Search noise. Numerous results returned "687" matches that are unrelated — e.g., a '687 patent in a Nitek v. Photon Wave LED case (E.D. Tex. 2:23-cv-00420), and paragraph "687" citations in unrelated pleadings (OptumInsight, Harbour Antibodies/Teneobio, ACQIS, etc.). None of these involve US 10,402,687.

3. Party/assignee details matter. Current assignee is Perceptive Automata LLC (Delaware entity Perceptive Automata, Inc. → name-corrected to Perceptive Automata, Inc. in 2019 → assigned to Perceptive Automata LLC in 2025). Security interests were recorded to Avenue Venture Opportunities Fund, LP (2021) and Piccadilly Patent Funding LLC (2025). Any litigation standing analysis should account for these assignments.

4. Do not conflate the asserted patents with the '687. Any statement that "Perceptive sued Tesla on the '687 patent" would be inaccurate based on the record I retrieved. The correct statement is that Perceptive sued Tesla on five family-member patents that claim priority to the '687 patent.

5. Conflicting "current date" inputs. I was given two different current dates (2026-04-26 in the task and 2026-09-30 in the system context). Because a June 2026 docket entry appears in the retrieved data, the status of the stayed W.D. Tex. case (particularly the unresolved Rule 12(b)(6) motion to dismiss and any voluntary dismissal) should be re-verified against live PACER/CourtListener before being relied upon.

---

If you want, I can attempt a further targeted pull (e.g., the full W.D. Tex. docket 7:25-cv-00594 to confirm the final disposition, or a Darts-IP/family-level litigation report for family 64903265) to confirm whether any non-Tesla defendant was ever sued on the '687 patent specifically. I cannot confirm that from the data retrieved so far, and I will not represent that it happened without a source.

I'll verify the structured "PTAB proceedings on file" block against live sources before writing this up.

Let me run additional targeted checks for any proceeding specifically naming the '687 patent.

Let me confirm there is no reexam/CBM/PGR against the '687 and check the family IPRs' current status.

PTAB / AIA Trial Proceedings — US 10,402,687

Compliance note first: the canonical structured block in my instructions ("PTAB proceedings on file") reports no AIA trial proceedings for this patent as of the most recent ingest. I ran independent web verification (USPTO PTACTS petition documents, Unified Patents PTAB portal, Ex Parte / GreyB PTAB mirrors, CourtListener RECAP) and found nothing that displaces that result — no IPR, PGR, or CBM petition naming U.S. 10,402,687 as the challenged patent. Everything below therefore reports a zero count for the '687 itself, with a clearly-labeled adjacent section on the family-level IPRs, which the prior litigation section already flagged and which I do not repeat wholesale.

---

Proceedings overview

There are zero AIA trial proceedings on US 10,402,687 — 0 active, 0 claims invalidated, 0 claims sustained, 0 settled, 0 institution denials. The bottom line for a defendant is not "the patent is hardened" in the sense of having survived attacks; it is that the patent is entirely untested at the PTAB and no § 315(e) estoppel or § 315(b) time-bar currently constrains a new challenge to it. The five IPRs that exist (Tesla v. Perceptive Automata, IPR2025-01573 through -01577) target five continuation patents that claim priority to the '687 — they do not put a single claim of the '687 at issue, and an adverse FWD in those proceedings would not cancel, narrow, or estop anything with respect to the '687's own claims. That is a materially better defensive posture than a patent that has been through — and survived — two IPRs, and a materially more predictable one than a patent already stripped of its independent claims.

---

Proceedings on US 10,402,687

(None. There is no proceeding number to report, and per my operating constraints I will not supply one by inference.)

No proceeding on file — USPTO ODP / PTACTS

• Type: N/A
• Filed: N/A
• Status: N/A — the structured USPTO Open Data Portal list is empty for this patent
• Judge panel: N/A
• Petition grounds: N/A
• Institution decision: N/A
• Final Written Decision: N/A
• Settlement / termination: N/A
• Appeal: N/A
• Defensive value: Two consequences follow directly from the emptiness. (1) No claim-level cancellation risk has been realized — claims 1 through the last claim of the '687 remain presumptively valid before the Board, and any district-court § 282 burden remains "clear and convincing." (2) No estoppel helps you and none hurts you: because no FWD has issued on the '687, § 315(e)(2) estoppel has not attached to anyone with respect to this patent, so the full universe of § 102/§ 103 patents-and-printed-publications grounds remains available to a first petitioner.

Caveat on confidence: I could not reach a live PTACTS "Post-Grant Review/Review Proceedings for U.S. Patent 10,402,687" landing page in this session. The absence is corroborated by multiple independent aggregators (none of which lists a proceeding against this patent number), but a defendant should re-confirm on PTAB E2E / PTACTS and via the PTAB's Patent Number search before relying on it in a filing.

---

Adjacent family-level proceedings (NOT on the '687 — read this as context, not as an attack on your patent)

These are the only AIA proceedings in the family. Every one of them challenges a different patent number. Tesla filed five IPRs between 2025-09-30 and 2025-10-10, challenging every claim of every patent Perceptive asserted in Perceptive Automata LLC v. Tesla, Inc.:

(Claim ranges are taken from Tesla's Notice of Broadened Sotera Plus Stipulation filed in the E.D. Tex. case; institution dates from the Unified Patents PTAB portal and Ex Parte PTAB case pages. Note one discrepancy in the record worth knowing about: Tesla's own Sotera notice lists the IPR2025-01573 petition date as September 20, 2025, while the proceeding-level data shows September 30, 2025. Treat the petition date for -1573 as 2025-09-30 pending confirmation against PTACTS.)

Why the '687 matters to these proceedings but is not in them: Perceptive's Patent Owner Briefs in IPR2025-01573 and IPR2025-01574 both recite that the challenged patents "claim[] priority to U.S. Patent Application No. 15/830,549, filed on December 4, 2017 (and which issued as U.S. Patent No. 10,402,687), as well as to U.S. Provisional Patent Application No. 62/528,771." That priority chain is the only reason the '687's number appears in the PTAB record. Filed 2017-07-05 (provisional) / 2017-12-04 (non-provisional) and granted 2019-09-03, the '687 is the family's earliest non-provisional and its shared-specification ancestor.

Grounds asserted in the family IPRs, for defensive-theory borrowing (not estoppel): Tesla's petitions lean heavily on Cox (WO 2014/210334) — a reference sharing a named inventor with the Perceptive family — combined with Ross (U.S. 10,496,091), and variously Huval (US 2018/0373980), Ogale (U.S. 10,733,506), Munro (US 2016/0162456), Zhang (US 2019/0072966), Djuric (US 2019/0049970), Zhu (US 2019/0164018), and Ellenbogen (US 2017/0099200), plus textbook excerpts (Probabilistic Robotics; Robot Motion Planning; Principles of Robot Motion). Petitioner's theme is that the '687's inventor failed to disclose Cox — his own prior PhD work — during prosecution, which Tesla characterizes in IPR2025-01577 as "the Patent Office's first opportunity to address prior art of a named inventor that was never submitted or considered during prosecution." A Cox-based obviousness theory is the most obvious candidate ground against the '687, given the shared specification and common inventor. If you are a defendant holding a demand letter on the '687, that is where I would start.

Procedural posture of the family IPRs as of the record retrieved: all appear past institution and into trial. The IPR2025-01577 docket shows Patent Owner Response activity around 2026-07-09, corrected PO Response 2026-07-09, Petitioner objections 2026-07-15, and a notice of deposition of Patent Owner's expert 2026-07-29. Perceptive filed a "Patent Owner's Brief in Support of Discretionary Denial" in each, and the Director issued a "Director Discretionary Decision: Refer" in IPR2025-01577 on 2026-02-03 (i.e., referral to a merits panel rather than outright discretionary denial) — a meaningful procedural event in the current Director-review regime, but again, one occurring in a proceeding on a different patent. No Final Written Decision has issued in any of the five. If instituted, statutory FWDs would be due roughly March 2027 (one year from institution, 35 U.S.C. § 316(a)(11)). Judge panels are not public on the sources I could reach — the case pages show "Pending Judge Assignment" — so I decline to name APJs.

---

Strategic summary

Claim status on the '687. Every claim of US 10,402,687 is UNTESTED — no claim is canceled and no claim is PTAB-sustained. There is no narrowing certificate, no adverse judgment, and no disclaimer of record. By contrast, the five family continuations all have every claim under challenge in an instituted or pending IPR. This asymmetry is the single most important fact for a defendant: if a demand letter you received cites the '687, that patent has never been before the Board, and the Board's eventual rulings on the continuations will not bind, cancel, or estop anything as to the '687's claims. Because the '687 shares its specification with the challenged continuations, however, a loss on Cox-based obviousness theories in the family IPRs could hand a future petitioner a ready-made roadmap against the '687, while a win for Perceptive on Cox would be persuasive (though not preclusive) in the other direction.

Estoppel landscape. Section 315(e)(2) estoppel is patent-specific. Tesla stipulated (a "Sotera Plus" stipulation) that if the PTAB institutes, Tesla will not assert at trial that any claim of the instituted patent is invalid on the instituted grounds, on grounds that reasonably could have been raised under §§ 102/103 on patents or printed publications, or on system-art combinations with the asserted references. That stipulation runs to the five asserted continuations — it creates no estoppel whatsoever against the '687, which Tesla never petitioned on. Likewise, § 315(b)'s one-year bar has not been triggered for the '687: Tesla was served with a complaint asserting five other patents, and a complaint that does not allege infringement of the '687 does not start the § 315(b) clock for the '687. Net: for the '687 specifically, the prior-art field is wide open and the time bar has not begun to run for any party who has not been sued on it. The practical constraint is the opposite one — the December 2025/2026 PTO rulemaking environment, in which the Director has tightened discretion, expanded "settled expectations" denials for older patents in force, and proposed § 42.108(d)–(f) bars. A patent that issued 2019-09-03 and expires 2037-12-04 is not ancient by those standards, but a '687 petition filed today would face the same discretionary-dental gauntlet that Perceptive invoked unsuccessfully so far here.

Pattern signals. One petitioner (Tesla) filed five IPRs against this family in a 10-day window — a coordinated, all-claims-cleared strategy tied to the district-court case, not a serial-attack pattern on a single patent. No defensive aggregator (Unified Patents, RPX, etc.) appears as a petitioner or RPI in this chain; Unified's portal merely hosts case data. Perceptive's conduct is instructive: rather than sit on the merits in the family IPRs, it aggressively pursued both discretionary denial and a stay-fight in district court, then voluntarily dismissed the E.D. Tex. case on 2025-12-29 and refiled in W.D. Tex. the same day — a venue-driven reset that also reset the § 315(b) analysis for the refiled case. Expect the same litigious-aggressive posture from this patent owner if the '687 is ever asserted.

Contradiction check with prior sections: none. The earlier litigation section's statement that "the '687 patent is not among the patents challenged in any of these IPRs" is confirmed by the underlying Sotera stipulation table and by each petition's Case Identification block. The earlier section's note that IPR2025-01573 and -01574 were "instituted" is consistent with the 2026-03-17 institution entries I found.

---

Recommended next steps

1. If you hold a demand letter citing the '687, do not treat the family IPRs as a shield or as a threat. Verify on PTAB E2E / PTACTS (patent-number search for 10,402,687) that no petition has been filed since the last ingest — the window is genuinely open, and being first-to-file carries real value here because there is no § 315(e) estoppel against the '687 and no § 315(b) bar running.
2. Mine the Tesla petitions for a ready petition skeleton. The Cox-based obviousness grounds (Cox alone, and Cox + Ross) as applied to the shared specification are the highest-probability attack on the '687. Start from the IPR2025-01573 petition (U.S. 10,614,344; claims 1–22) and IPR2025-01577 ('346) as the closest claim-set analogues, and confirm whether the '687's claim set adds limitations not addressed in those grounds.
3. Watch the family FWDs as a leading indicator. FWDs in IPR2025-01573 through -01577 are due approximately March 2027 under § 316(a)(11); institution decisions for -01576 and -01577 were expected around 2026-04-16 per Perceptive's 2026-02-20 filing, with -1575 and -1577 showing institution on 2026-03-23. The first FWD to issue will tell you whether the Board buys Cox as § 103 art against this family — which is the single most useful data point for pricing a '687 challenge.
4. Do not cite a '687 FWD or IPR number in any pleading. There isn't one. If opposing counsel represents otherwise, the correct response is that the '687 — priority document of the asserted patents — has never itself been challenged at the PTAB, and that the family IPRs create no estoppel and no cancellation effect as to its claims.
5. Re-verify before filing. My verification relied on aggregators (Unified Patents, Ex Parte, GreyB/IPVerse) plus USPTO PTACTS petition documents rather than a live PTACTS proceeding-list page for the '687, and one petition date in the record (-1573) is internally inconsistent. Confirm the zero count and the institution dates against PTACTS and the W.D. Tex. docket 7:25-cv-00594 on CourtListener before relying on any date here.

I'll research the assignment record for US 10,402,687 across the USPTO Assignment Center, Google Patents legal events, and NPE directories.

Let me try more targeted sources that index the USPTO reel/frame data and correspondents.

Ownership Reconstruction — US 10,402,687

Data-source note (read first): The authoritative facts below come from the Google Patents legal-events record in the patent text supplied for this analysis (which reproduces the USPTO reassignment entries), plus the prior sections. I attempted to pull the full USPTO Assignment Center records (reel/frame, correspondent of record) via web search; the Assignment Center's per-patent record did not surface through the search results I could reach, and I hit the tool-call limit before completing a live lookup. Only one reel/frame number is independently confirmed (reel 045672/0956, referenced inside the corrective-assignment entry itself). Everything else is dated but without a verified reel/frame or correspondent. I have flagged each item accordingly rather than inventing reel/frame pairs.

---

Inventors

Pattern note: All three inventors assigned their entire interest to the company within ~4 months of the Dec 4, 2017 filing (recorded Apr 30, 2018) — i.e., they were founders/employees, not outside licensors. This is not the "inventors bolt within 12 months" fire-sale tell; it is the ordinary founder-assignment pattern. What is notable is what happened later to the assignee entity, covered below. Employer at filing is determinable only as "Perceptive Automata, Inc." (the assignee); no separate employer is documented in the sources I could reach.

---

Original assignee

Perceptive Automata, Inc. (originally recorded simply as "Perceptive Automata"; name corrected by the 2019 corrective assignment to Perceptive Automata, Inc.).

• Business: Venture-backed autonomous-vehicle software startup spun out of Harvard research; its product was a vision/AI system that predicts pedestrian and road-user intent for self-driving cars — i.e., it practiced the very subject matter claimed (the specification describes the product pipeline in FIGS. 3, 9, 10).
• Shipped a product? Yes during its operating life — the patent describes an in-vehicle sensor → trained-model → intent-prediction module (see FIG. 10, "real-time module... installed on a car"). I could not verify current commercial availability.
• Current status: The original Perceptive Automata, Inc. apparently ceased operations, and its IP was transferred to Perceptive Automata LLC (a Texas LLC formed in 2025) per the Feb 19, 2025 assignment. The LLC is the plaintiff now asserting the continuation patents against Tesla. Caution: the exact wind-down mechanism (asset sale vs. assignment in lieu vs. bankruptcy) is not established by anything I retrieved; I will not characterize it as a bankruptcy fire-sale.

---

Assignment timeline

Reel/frame confirmed only where stated. Where the record shows a recording event without a reel/frame in the source text, I say "reel/frame not retrieved" — these should be pulled from Assignment Center before being relied on.

• Recorded 2018-04-30 — Reel 045672/0956 (confirmed: this is the frame the 2019 corrective assignment expressly corrects)

  • Conveyance: Assignment of Assignors' Interest (initial, all three inventors)
  • Assignor: Samuel English Anthony; Kshitij Misra; Avery Wagner Faller
  • Assignee: Perceptive Automata (name later corrected)
  • Correspondent: not retrieved — not verified
  • Context: Founder/employee assignment of invention to the company (acquisition of inventors' rights, standard).

• Recorded 2019-05-23 — Reel/frame not retrieved

  • Conveyance: Corrective Assignment — "CORRECTIVE ASSIGNMENT TO CORRECT THE ASSIGNEE'S NAME PREVIOUSLY RECORDED AT REEL: 045672 FRAME: 0956"
  • Assignor: Anthony; Faller; Misra (same inventors)
  • Assignee: Perceptive Automata, Inc.
  • Correspondent: not retrieved — not verified
  • Context: Change of name only (entity designation cleanup; no ownership change).

• Recorded 2021-04-01 — Reel/frame not retrieved

  • Conveyance: Security Agreement
  • Assignor: Perceptive Automata, Inc.
  • Assignee: Avenue Venture Opportunities Fund, LP (as secured party)
  • Correspondent: not retrieved — not verified
  • Context: Securitization — venture-debt lender takes a security interest in the portfolio. This is collateral, not an ownership transfer.

• Recorded 2025-02-19 — Reel/frame not retrieved

  • Conveyance: Assignment of Assignors' Interest
  • Assignor: Perceptive Automata, Inc.
  • Assignee: Perceptive Automata LLC
  • Correspondent: not retrieved — not verified
  • Context: Transfer to successor/assertion entity — near-identical name, new LLC formed 2025; this is the link where the operating entity hands the family to the litigating entity.

• Recorded 2025-03-25 — Reel/frame not retrieved

  • Conveyance: Patent Security Agreement
  • Assignor: Perceptive Automata LLC
  • Assignee: Piccadilly Patent Funding LLC, as Security Holder
  • Correspondent: not retrieved — not verified
  • Context: Securitization / patent litigation funding — a funding vehicle takes a collateral interest in the very portfolio being asserted; ~1 month after the LLC acquisition and ~4 months before the first Tesla complaint.

Important limitation: Assignment Center does carry the correspondent of record for each of these, and that is the single most diagnostic field for the "repeat attorney behind re-named LLCs" test. I could not retrieve those correspondents from the available search results. This section should be treated as complete on dates and parties but incomplete on correspondents. Verify at the USPTO Assignment Center: https://assignmentcenter.uspto.gov/ (mirror: https://assignment.uspto.gov/patent/index.html) by searching patent number 10402687.

---

Timeline diagram

timeline
    title Ownership of US 10402687
    2017 : Provisional filed Jul 5
         : Nonprovisional filed Dec 4
    2018 : Inventors assign to Perceptive Automata
    2019 : Patent issued Sep 3
         : Corrective assignment fixes assignee name
    2021 : Security agreement to Avenue Venture
    2025 : Assigned to Perceptive Automata LLC
         : Security agreement to Piccadilly Patent Funding
         : Tesla suits filed on continuation patents

---

NPE / troll-pattern signals

1. Shell-entity transfer — unclear. The Feb 19, 2025 transfer is not to a classic "…IP Holdings/Licensing" shell but to Perceptive Automata LLC, essentially the same name as the assignor, which is more consistent with a corporate restructuring/asset transfer than with a purpose-built anonymous shell. However, the LLC was formed in 2025 and (per the prior sections) the original Inc. ceased operations — which is the functional equivalent of a non-practicing successor. I cannot confirm the LLC's address or registered-agent status from the retrieved data, so I mark this unclear, not present. Pull the LLC's formation/registered-agent record to resolve.

2. Known asserter in the chain — not present (on the named lists). Neither Perceptive Automata LLC nor Piccadilly Patent Funding LLC appears among the classic directory names given (Acacia, Marathon, IV, IPNav, Wi-LAN, Conversant, Vringo, Pendrell, Round Rock, etc.) in anything I retrieved. Caveat: Piccadilly Patent Funding LLC is, by its name and the "as Security Holder" capacity, a patent-finance entity; I could not confirm an RPX/Unified high-frequency-plaintiff listing for it. Do not treat "not present" as exculpatory — it means "not matched."

3. Repeat correspondent across the chain — unclear (data not retrieved). This is the key gap. There are five recorded events, and if the same attorney/firm filed the 2018 founder assignment, the 2021 Avenue security agreement, the 2025 LLC assignment, and the 2025 Piccadilly security agreement, that would be a strong recurrence finding. I could not retrieve any correspondent name, so I cannot make the call either way. Recommend a direct Assignment Center pull.

4. Cascading transfers — present (moderate). Two ownership/security events in ~5 weeks: assignment to Perceptive Automata LLC (2025-02-19) immediately followed by a patent security agreement to Piccadilly Patent Funding LLC (2025-03-25). Layered on the earlier Avenue (2021) security interest, the portfolio shows a stacked-collateral structure characteristic of monetization financing. Not a multi-LLC cascade, but a compressed sequence.

5. Pre-litigation transfer — present (with a date caveat). The transfer to the asserting Perceptive Automata LLC (2025-02-19) precedes the first Tesla complaint. Timing depends on which filing date is correct — see the contradiction flagged below. If the E.D. Tex. suit was filed 2025-07-23, the gap is 5 months → within the 6-month window (signal present). If the earliest suit was 2025-12-29, the gap is ~10 months → outside the window (signal weak/absent).

6. Bankruptcy fire-sale — unclear / not established. The Inc. "apparently ceased operations," but no Chapter 7/11 filing or court-supervised patent sale is documented in the sources I reached. Do not assert a fire-sale.

7. Privateering — not present. Privateering requires an operating parent transferring to an NPE to assert against competitors. Here the operating company is itself gone and its own successor is the plaintiff; there is no separate manufacturer behind the assertion.

8. Defensive aggregator — not present. The chain does not terminate at RPX, AST, LOT, Unified, or OIN. The Piccadilly security interest is a monetization/financing structure, the opposite of a defensive neutralization.

Contradiction to flag

The two prior sections disagree on the first Tesla filing: the patent summary states the earlier E.D. Tex. case 2:25-cv-00742 was filed Dec. 29, 2025; the litigation summary states it was filed July 23, 2025. This is a 5-month discrepancy and it directly determines signal #5. It should be resolved against PACER/live docket before relying on the pre-litigation-transfer finding.

Scope caveat that cuts against over-reading this chain

Everything after the 2019 corrective assignment is a portfolio-level event (family-wide security agreements and the LLC transfer). The asserted patents are the continuations (10,614,344; 11,126,889; 11,467,579; 11,520,346; 11,753,046), not the '687 itself. So these signals attach to the '687's current owner, but the assertion behavior they evidence is not directed at the '687. Any troll/NPE characterization is about the entity and the portfolio, not about litigation over this specific patent — which has not been asserted.

---

Verdict

NPE — moderate confidence.

The '687 was developed by a genuine operating company that practiced the claims (Perceptive Automata, Inc.), but that entity ceased operations and, on 2025-02-19, assigned the portfolio to Perceptive Automata LLC — a 2025-formed successor that (per the prior sections) does not ship products and is now asserting the family against Tesla — followed 2025-03-25 by a patent security agreement to Piccadilly Patent Funding LLC, as security holder. Two weak signals are present (compressed cascading transfers, and a likely pre-litigation transfer, subject to the filing-date contradiction), and the current holder is a non-practicing entity with third-party litigation funding — but no confirmed shell-entity name, no confirmed known-asserter match, and no retrieved correspondent of record prevent a high-confidence call. This verdict is provisional pending (a) the Assignment Center correspondents of record and (b) resolution of the Dec 29 vs. Jul 23, 2025 filing date.

Verify at: USPTO Assignment Center — https://assignmentcenter.uspto.gov/ (search patent 10402687); mirror https://assignment.uspto.gov/patent/index.html. I could not independently retrieve the reel/frame set or correspondents and will not represent them as confirmed.

I'll research the cited prior art for US 10,402,687. Let me pull the patent's citation records from multiple sources.

Let me pull the complete citation list for the '687 patent and the examiner's cited references.

Let me retrieve the granted patent's front-page "References Cited" list directly from the patent PDF and Justia's citation listing.

Scope and method note (read this first)

I searched for the literal grant number 10402687 / US10402687B2. The results returned the correct patent (Perceptive Automata, "System and method of predicting human interaction with vehicles," granted 2019‑09‑03, App. 15/830,549). I did not substitute or auto-correct to any similar number, and I discarded look‑alike hits (e.g., US 10,466,487 to Pogotec; US 10,464,976; the Nitek "‘687" docket noise).

Coverage caveat — important: I was able to retrieve only a partial set of the cited references. Google Patents shows the '687 with "Citations (11)", but the retrievable text truncated the table after the seventh entry, and I could not pull the separate "Non‑Patent Citations" block or the front‑page PTO‑892 (the search tool limit was reached). So the list below is incomplete and must be verified against the printed patent front page before it is relied on. I have flagged exactly where the record is truncated rather than filling the gap.

---

Confirmed cited patent references (as listed on the '687 patent record)

Not retrieved (of the 11 total): the remaining ~4 patent citations and all non‑patent citations. Do not treat the above as the complete set.

Field observation: Every confirmed citation is pre‑2008 surveillance / video‑analytics / human‑activity‑monitoring art. That is consistent with the examiner having used these as background references (the sort that get category "A" — "general state of the art not of particular relevance") rather than as the anticipatory core of the rejection.

---

§ 102 analysis — reference by reference

Controlling date: the '687's effective filing date is Jul. 5, 2017 (provisional 62/528,771), with the non‑provisional filed Dec. 4, 2017. All seven references above published years earlier, so each qualifies as prior art under 35 U.S.C. § 102(a)(1) as a printed publication. Qualification is not the issue; disclosure is.

Claim 1 (method), claim 11 (system), and the likely further independent claim require, in substance: (a) receiving image/video of a road scene from the perspective of a participant in that scene, the scene including a pedestrian/cyclist/vehicle; (b) generating stimulus data that may be an altered version of the image/video; (c) transmitting the stimulus data to a user interface; (d) receiving plural response data of human‑inputted action / likelihood‑of‑action judgments; (e) aggregating a subset of responses into statistical data; (f) creating a model based on the statistical data; (g) applying it to a second image/video; (h) generating a behavior prediction.

None of the seven confirmed references discloses element (d)–(f): crowd‑sourced human behavioral judgments on altered driving-perspective stimuli aggregated into a statistical training target, and a model trained on that target. That is the patent's asserted point of novelty, and it is absent from surveillance art, which generates machine predictions directly rather than learning from aggregated human observer judgments.

Consequently:

• No confirmed citation anticipates claim 1, claim 11, or the remaining independent claim under § 102. Anticipation requires a single reference disclosing every element arranged as claimed (Net MoneyIN v. VeriSign); none of these comes close to elements (c)–(f).
• Closest individual references, and the limitations they could arguably touch:
  • US 2005/0265582 A1 (Buehler) and US 2007/0182818 A1 (Buehler) — tracking/monitoring behavior of multiple objects. These are the only confirmed references that even approach "(h) generate a prediction of behavior." They do not disclose ego‑perspective road‑scene capture, crowd‑sourced human judgment capture, response aggregation, or model training on aggregated statistics. Under § 102 they fail; at most they are § 103 combination art against the detection/tracking limitations of claims 1 and 11 (and the corresponding limitations in the dependent claims).
  • US 2003/0058111 A1 (Philips) — vision‑based monitoring of a human's state. Potentially relevant only as background to the "state of mind" concept in the specification; discloses no stimuli, user interface, or aggregated human labels.
  • US 2006/0279630 A1 (Aggarwal) — situational awareness by data fusion. Background only; no human‑in‑the‑loop labeling pipeline.
  • US 5,909,548 (Sensormatic) and US 2004/0161133 A1 (Elazar) — operator alerting / alarm management. General video‑analytics background; disclose no modeling‑from‑human‑judgment step.
  • US 2007/0291117 A1 — I cannot responsibly map it to a claim without its title and text.

Which claims are potentially implicated (best case for a § 102 challenge, still failing): at most the pre‑characterization limitations of independent claims 1 and 11 — "receives an image and/or video segment ... including at least one of a pedestrian, a cyclist, and a motor vehicle" and the notion of predicting behavior. Even there, the references do not disclose the "perspective of a participant in the road scene" limitation, which is where an anticipatory reference would have to start.

Dependents: none of the confirmed references supplies the dependent‑claim subject matter either — e.g., cl. 2 (supervised learners: random forest / SVR / CNN / RNN / LSTM), cl. 3–4 (pixel manipulation; looped frames), cl. 5 (ordinal likelihood values), cl. 6–7 (response time and eye‑position parameters; central tendency/variance/skew/kurtosis/scale/histogram), cl. 10 (error‑signal weight adjustment). US 2005/0265582 and US 2007/0182818 offer generic tracking, not ordinal human‑judgment aggregation.

---

Where the real § 102 risk lies (and it is not on the face of the '687)

Two things you should verify, because they are more dangerous than the face citations:

1. Non‑patent citations. The '687 record has a non‑patent‑citation block I could not retrieve. Given the claim language (crowd‑sourcing via Amazon Mechanical Turk is named in the spec, and the "Likert scale"/ribbon‑judgment mechanisms), the most plausible § 102 exposures are human‑computation / crowd‑labeling publications and pedestrian‑intent‑prediction papers, not the surveillance patents. These need to be pulled from the front page.

2. WO 2014/210334 (Cox et al.). The earlier litigation research (already generated) established that Tesla's IPR campaign against the continuation patents relies on WO 2014/210334, a reference naming Perceptive founder Samuel Anthony. It is not among the confirmed face citations of the '687. If that reference has an effective date before Jul. 5, 2017, it is a materially stronger § 102/§ 103 reference against this family than anything listed above, and its applicability to the '687 specifically should be checked against the IPR claim charts rather than assumed. I am flagging this as a lead, not as a verified '687 citation.

---

Confidence and limitations

• High confidence: the identity of the '687; the citation entries #1–#6 (numbers, assignees, priority/publication dates, titles) as returned from the patent record; and the conclusion that none of the confirmed references anticipates claims 1 or 11 under § 102.
• Moderate confidence: that the confirmed set comprises examiner‑cited background art; the category labels ("A") are my inference from the field and vintage, not read off a PTO‑892.
• Not established / do not rely on without verification: the identity of citations #7's title and of the ~4 unretrieved citations; the non‑patent citations; whether any of the 11 is categorized "X" or "Y" on a foreign search report; and the applicability of WO 2014/210334 to the '687.
• Recommended next pull: the US10402687B2 front page PDF at patentimages.storage.googleapis.com (PTO‑892 + Non‑Patent Citations), the file wrapper on USPTO PatentCenter for App. 15/830,549 (to separate examiner‑cited from applicant‑cited references), and the EP3701418 / WO2019112912 search report, which often carries X/Y categorizations that the US face does not.

Sources used: patents.google.com/patent/US10402687B2/en (citation table, abstract, family); the Google Patents "Patent Citations (11)" table as returned in search; wiki.golden.com record for US Patent 10402687 (bibliographic cross‑check).

I'll research the prior art landscape for this patent before analyzing obviousness.

Let me pull the specific prior-art references Tesla relies on across the family IPRs, since those are the closest art to the '687.

Let me verify bibliographic dates for the key references, since §103 prior-art status relative to the '687's July 5, 2017 priority date is decisive.

Obviousness Analysis of U.S. Patent No. 10,402,687 Under 35 U.S.C. § 103

Important scoping note before the analysis begins

On "the Prior Art section of this page": The Google Patents page you supplied was fetched in truncated form — it captures the bibliographic field, the Definitions, the Description and the Abstract, but the fetch stops before the front-page "Citations / Cited By / Similar Documents" listing. I therefore could not read the page's own prior-art citation table, and I will not invent its contents.

What follows is grounded in (a) the prior-art references actually in the record for this patent family — the references Tesla, Inc. papers in its co-pending IPRs against the '687's continuation/priority descendants, which share the '687's specification and, for the independent claims at least, its disclosure — and (b) the admissions in the '687's own specification and Background, which are the most reliable art-of-record for the '687 itself because they are the patent's own words. Grounds (b) is genuinely available to a challenger: the '687's Background expressly characterizes motion-vector extrapolation as the prior approach, and its specification expressly characterizes the machine-learning algorithms it uses as a menu of known options.

Also flagging a status asymmetry. No IPR, PGR, or district-court assertion naming US 10,402,687 has been located (consistent with the earlier sections of this analysis). Every institution decision I found — IPR2025-01573 (instituted 03/17/2026), IPR2025-01574 (instituted 03/17/2026), IPR2025-01575 (instituted 03/23/2026) — concerns the descendants (US 10,614,344; 11,126,889; 11,753,046). So this §103 analysis is prospective, and its force depends on an inverse relationship worth stating up front:

The '687 is the broadest member of the family. Its independent claims do not recite autonomous-vehicle control, do not recite "state of mind" as a required output, and do not recite a "training dataset comprising summary statistics of user responses." Per the earlier-generated claim summary, claim 1 of the '687 ends at "generates a prediction of user behavior in the second image or video segment." The continuations add those limitations. If Tesla is prevailing (or likely to prevail) against the narrower descendants on the Ross/Cox/Huval/Munro/Djuric art, the broader '687 claims are, a fortiori, exposed to the same art — subject only to the priority-date question discussed in §2.

---

1. Governing framework

Statute: AIA 35 U.S.C. § 103. The '687 was filed Dec. 4, 2017 (App. 15/830,549) with priority to provisional 62/528,771, filed July 5, 2017. It is therefore an AIA patent, and § 102(a)(1)/(a)(2) with the § 102(b)(1)/(b)(2) exceptions govern prior-art status. I use July 5, 2017 as the critical date throughout (and flag in §2 where a reference's status turns on it).

Standard: Graham v. John Deere Co., 383 U.S. 1 (1966) (scope/content of art; differences; PHOSITA level; secondary considerations) and KSR Int'l Co. v. Teleflex Inc., 550 U.S. 398 (2007) (a combination is obvious where a POSITA had a reason to combine, including the rationales Tesla invokes — known technique improving similar devices in the same way; predictable use of prior-art elements; design incentive from a finite set of identified options).

PHOSITA (proposed, as of July 5, 2017): A person with at least a bachelor's degree in computer science, electrical engineering, robotics, or a related field, and roughly two to three years of experience (or a master's with less), in computer vision and/or machine learning applied to vehicle perception or driver-assistance systems. This is the level Tesla's petitions appear to assume and Perceptive's POPRs do not appear to contest as such. I have not verified Perceptive's proposed alternative level.

Claim construction of record: The '687's specification and file history are the only construction sources I can verify. Two terms matter:

• "stimulus data … comprising at least one of the [image/video], and an altered version" — importantly, this is in the alternative. The unmodified footage alone satisfies the limitation. This defeats any non-obviousness argument that depends on the manipulation feature being required. (Manipulation is separately recited in dependent claims 3–4, which is where it must be attacked.)
• "prediction of user behavior" — on the '687's face this is behavior, not "state of mind." The '889/'344 family members recite the state-of-mind output; the '687 does not. This is the single most consequential scope difference.

Printed-matter overlay: In the family IPRs Tesla argues that limitations reciting stored/aggregate informational content (the "statistical data" and parameter recitations) are printed matter lacking a functional relationship to the hardware, and therefore are entitled to no patentable weight for §103 purposes. Perceptive's POPRs counter that "The Challenged Claims Are Entitled To Patentable Weight." This dispute is directly transferable to the '687, and if Tesla prevails on it the '687's aggregation and parameter limitations (claims 1[e]–[f], 6–7) cannot be relied on to distinguish the art — a materially easier §103 case. I treat it below as an alternate theory, not a premise.

---

2. The prior-art references, with dates and confidence

The '687's own admissions — the strongest and least assailable art. The Background states that autonomous vehicles "may rely on methods that make decisions on how to control the vehicles by predicting 'motion vectors'" and that these methods "do not predict a person's actions or movements based on other observations besides his or her current and past movements." That is a conceded state of the art, and it is the exact starting point for a motivation-to-improve analysis.

One inventor-overlap caveat to carry forward. The earlier sections report that Tesla's IPR grounds rely on WO2014/210334 (Cox) and that Cox is said to name Samuel Anthony, a named inventor of the '687, and that the '889 petition describes Cox as a reference "which shared an inventor with the '889 patent." If that is accurate:

• It creates a potential patent-owner argument that Cox is a §102(b)(2)(C) commonly-attributed disclosure. But that exception does not save the '687. Cox published in December 2014 — more than a year before the July 5, 2017 critical date — so Cox is §102(a)(1) art, and §102(b)(1) requires the excepted disclosure to have been made one year or less before the effective filing date. The grace period is unavailable on the face of the dates. The inventor overlap is therefore immaterial to §103 and instead becomes a duty-of-disclosure/inequitable-conduct fact — which is consistent with the earlier section's note that Tesla "has already indicated in its initial disclosures that it intends to pursue inequitable conduct claims." That is a validity exposure for the '687 that is independent of §103, and it is worth flagging because the '687 is the application in which the alleged non-disclosure would have occurred.

---

3. The claim 1 limitations to be met

Using the limitation structure from the earlier-generated claim summary (claim 1, "A computerized method of predicting user interaction with vehicles"), cross-checked against Justia's listing:

Claims 11 (system) and the expected ~claim 15 (non-transitory CRM — unverified, per the earlier section) recite the same sequence with no additional substantive limitation, so each ground below applies to all independent claims with the substitution of apparatus/CRM language.

---

4. Ground 1 — Cox in view of Ross (and Huval as the annotator-portal teaching)

Primary combination: Cox (WO 2014/210334) + Ross (US 10,496,091). Secondary: Huval (US 2018/0373980) for the crowdsourced-annotation hardware/UI path.

Element mapping:

• [1.a] — Ross discloses perception data from an autonomous vehicle's sensors for road users who are pedestrians, bicyclists and vehicles (US 10,496,091, Abstract and 15:1–16:52). Huval discloses images captured by vehicle cameras of "vehicles, pedestrians, and cyclists." The "perspective of a participant in the road scene" is met by the ego-vehicle-mounted camera in either.
• [1.b] — Met in the alternative by the unaltered image/video itself. Neither Cox nor Huval needs to manipulate anything for this limitation; each transmits captured imagery to human viewers/annotators. (Manipulation is pushed to claims 3–4.)
• [1.c] — Huval is the cleanest teaching: images are stored remotely and "queue[d] … for transmission to human annotators for manual labeling," rendered "in an annotator portal" that "execute[s] on a desktop computer" — i.e., transmission to a user interface. Cox separately discloses presentation of images to a human population for judgment.
• [1.d] — Cox: human observers supply judgments; Huval: annotators place bounding boxes and select object-type labels for pedestrians, cyclists and vehicles. Ross supplies the action/likelihood content: the intent model outputs "possible next actions (and predicted trajectories), a given point in time when the action is likely to occur, and associated likelihood values," where the actions come from a "predetermined list of actions" (Ross, 1:55–2:13; 4:33–65; 9:13–10:10).
• [1.e] — Cox is the key reference: it teaches compiling aggregate statistical information from responses across a human population, including psychometric data such as response times and accuracy, and using that statistical data as the training target. This is the '687's "aggregate a subset of the plurality of response data … to form statistical data," including the "content of a response" and "time associated with entering a response" parameters of dependent claim 6.
• [1.f] — Cox teaches training a supervised model on that statistical data so the model "better mimic[s] human performance."
• [1.g]–[1.h] — Ross teaches executing the trained model on new perception data to generate a prediction of the road user's next action and trajectory and then maneuvering on it. The application to a "second" image/video is the ordinary deployment step Ross describes.

Motivation to combine (articulated reasons):

1. Cox expressly directs its technique to the same field. Cox suggests applying human-population-trained models to "driverless" automobile applications. A POSITA reading Ross's stated problem — that plain behavior prediction "fail[s] to consider what that pedestrian wants to do or rather, his or her intent" (US 10,496,091, Background) — has an express reason to improve the model's outputs to reflect human judgment.
2. KSR rationale (a): known technique improving a similar device in the same way. Cox's human-judgment-aggregation is a known training technique; Ross's intent model is the same class of device (a supervised/discriminative road-user predictor). Applying the former to the latter predictably yields a more "human-like" predictor — which is the '687's own stated objective ("in a way that more closely resembles how human drivers would judge the behavior").
3. KSR rationale (b): predictable result. No change in the underlying hardware, sensor suite, or network architecture is required. The data pipeline (vehicle images → annotations → statistics → trained model → deployed model) is identical in topology to Cox's and Huval's; only the content of the human judgment changes (from object labels to intent/action likelihoods).
4. Reasonable expectation of success. Cox reports that population-derived statistics improve model fidelity; Ross provides the label space (next action + likelihood) to be populated by that process.

Anticipated counterargument. Perceptive's POPRs argue that a §103 ground is improper where the petition uses the challenged patent "as a blueprint" for the combination (a hindsight-bias objection), and that a petition must address the Graham factors and articulate the difference between the claims and the art. That objection is procedural and somewhat formalistic, but it has teeth at institution stage: it is the articulation of the difference between claim 1 and the references — not the existence of the references — that Tesla must supply. For the '687, the required articulation is easy and specific: neither Cox nor Ross, alone or together, expressly describes generating the training target from the distribution/aggregate of human predictions of the road user's future behavior, as opposed to from ground-truth observed behavior — which is why Grounds 2–3 exist.

---

5. Ground 2 — Cox + Ross + Munro

Add Munro (US 2016/0162456) for [1.e] specifically: Munro teaches tallying annotator responses into summary statistics — that is, converting a plurality of individual human responses on one item into an aggregated statistic.

Why this addition is not redundant (and why the distinction matters for the '687):

• [1.e] recites aggregating "a subset of the plurality of response data" — the word "subset" imports a selection/weighting step. The '687's specification supplies the disclosure for what the subset is selected on: excluding observers by "extrinsic or intrinsic criteria such as the time it took an observer to respond, the geographical location of an observer, the observer's self-reported driving experience, or the observer's reliability" (see also claim 6–7 and the FIG. 6 column 606 response-time weighting).
• Cox supplies exactly that selection mechanism — its psychometric parameters (response time, accuracy) are the filter criteria. So the combination Cox + Munro renders [1.e] obvious with a specific motivation: Munro's tally is the aggregation step, Cox's psychometrics is the "subset" selection criterion.
• Motivation: a POSITA seeking a robust training target would not rely on a single annotator; aggregating across annotators is the routine way to suppress label noise, and Ross itself needs a likelihood (a probability, hence an aggregate) for each candidate intent. Munro supplies the arithmetic; Cox supplies the target semantics.

Anticipated counterargument. Perceptive's POPR for the '889 argues that the Petitioner "failed to identify with particularity the material disclosure of the references where the recited claim language is allegedly met" and instead "cites to large sections of references and relies on innumerable string citations." A Ground 2 built on Cox + Ross + Munro is vulnerable to the same criticism unless the petition pins each limitation to a column/paragraph and a claim element. My mapping above does that; a real petition would need paragraph-level pin cites to Munro's tallying disclosure, which I have not independently verified.

---

6. Ground 3 — Huval + Ross + Cox (+ Munro)

This is the ground Tesla actually ran against the '889 (IPR2025-01574, Grounds 1–4: Huval-Ross; Huval-Ross-Cox; Huval-Ross-Munro; Huval-Ross-Cox-Munro). Its transfer to the '687 is straightforward because the '687's claim 1 omits the vehicle-control and "state of mind" limitations that the '889 adds:

• Huval supplies [1.a], [1.b], [1.c], [1.d] (the annotator portal, bounding boxes, object types).
• Ross supplies the action/likelihood label semantics and the deployment-on-new-data step for [1.d], [1.g], [1.h].
• Cox supplies [1.e], [1.f] (statistical aggregation of human responses; training a supervised model to mimic humans).
• Munro supplies the tallying for [1.e] if Cox's aggregation is found insufficiently express.

Motivation: Ross provides a use for Huval's labeled images beyond object detection — predicting intent — and Huval supplies a concrete, already-built pipeline for collecting the human labels that Cox says are needed. The combination is a three-way complementarity with no overlap in function: Huval = data collection at scale; Cox = how to turn human judgments into a training target; Ross = what to predict and what to do with the prediction. This is the classic KSR "familiar elements according to known methods … yielding predictable results" posture.

---

7. Ground 4 — Cox (or Huval) + Ogale (+ Ross)

Ogale (US 10,733,506) is offered only for the proposition that a model trained on "multiple human participants'" labels, when executed, outputs predictions that "predict[] how a human would have labeled the input sensor data."

For the '687 this is less necessary than for the continuations, because the '687's [1.h] requires only "a prediction of user behavior," not a prediction of a human's label. Ogale's value is therefore defensive: it forecloses an argument that the Cox-model's output must be characterized as a human-response distribution. Motivation: Ogale "merely makes explicit a desired outcome" of the Cox + Ross combination, providing a known technique with minimal modification.

---

8. Ground 5 — Djuric + AA + Ross (+ Cox)

From IPR2025-01576 (the '579), Tesla's Grounds 1–6 are Djuric-AA permutations with Cox, Rolfe and Ross. Djuric's goal-based object-movement prediction and AA's sequence model supply alternative trajectory-prediction machinery; Ross supplies intent/next-action prediction; Cox supplies the human-judgment training target.

For the '687 this ground is cumulative rather than necessary — Djuric/AA map to [1.h] (behavior prediction), which Ross already covers — and it is the weakest of the five grounds because it requires the most references to reach the same elements. It becomes useful only if a limitation is construed to require a goal-based/context-conditioned model rather than a generic supervised predictor. Given the '687's plain "applying the model" language, I would not expect that construction.

---

9. Dependent claims of the '687

---

10. Secondary considerations and the strongest patent-owner rebuttals

I found no secondary-considerations evidence in the record I could retrieve for the '687. But the '889 complaint (the earlier-cited authority list, Perceptive Automata LLC v. Tesla, Inc., 7:25-cv-00594) telegraphs the argument Perceptive will make, and it is the argument most likely to be recycled for the '687:

1. Nexus/commercial-success theory. Perceptive pleads that the claimed subject matter was not well-understood, routine or conventional as of July 2017, proffering (i) Tesla's own later-filed applications (e.g., US 2021/0271259; 2020/0348909; 2023/0176593; 2022/0284712; US 11,748,620; 11,816,585; 11,537,811; 11,215,999; 11,150,664; 11,636,333; 10,678,244) as admissions that a need existed, (ii) the fact that Waymo ran with safety drivers until Nov. 2017 and Tesla did not ship FSD beta until Oct. 2020, and (iii) contemporaneous 2017 literature (Zyner et al., "Long Short Term Memory for Driver Intent Prediction," IEEE IV, June 2017; Deakin University VRU-interaction work) stating that reproducing driver intuition "is still an open area of research."
   • This is a §101-eligibility argument ("not well-understood, routine, and conventional to a skilled artisan," Berkheimer) that Perceptive is repurposing for §103. It is a double-edged tactic. The 2017 literature Perceptive cites actually supports the §103 case at the level of the '687's independent claims: Zyner et al. (June 2017), cited by Perceptive itself, describes using LSTMs to infer driver intentions — i.e., the deep-learning intent-prediction architecture was already being published in the art weeks before the critical date. A POSITA reading that literature would have had a textbook motivation to apply the same sequence-model machinery to pedestrian/cyclist intent, which is precisely what Grounds 2–4 do.
   • Where Perceptive's point has real force is at the level of the specific combination recited in claim 1: the 2017 literature is about inferring intent from kinematics/context, not about building the training target from an aggregate of human observers' predictions. That is the genuine invention here, and it is why Cox is the load-bearing reference for the '687. If Cox is neutralized — or if a court holds that Cox does not fairly teach generating the target from aggregated human predictions rather than from aggregated human annotations — Grounds 1–3 collapse for claims 1, 11 and the CRM claim, and the case reduces to the printed-matter and AAPA theories.
2. Redundancy objection. Perceptive's POPRs repeatedly argue the multi-permutation grounds are redundant and that a petitioner must explain why one ground is preferred. Any §103 attack on the '687 should therefore be presented as one primary two-reference ground (Cox + Ross) with a single clearly-motivated third reference (Munro), plus a single fallback, rather than as five permutations.
3. Graham-factor articulation. Perceptive's POPRs fault the petitions for failing to articulate the difference between the claims and the art. For the '687 the difference to be articulated is narrow and must be stated precisely: the art must be shown to teach generating the training target from an aggregate of human observers' predictions of road-user behavior. Everything else in claim 1 is squarely in Cox/Huval/Ross/Munro.

---

11. Bottom line

• Claim 1 (and independent claims 11 and the CRM claim) of the '687 would have been obvious over Cox + Ross, with Munro supplying the response-tallying/aggregation step and Huval supplying the crowdsourced annotator-portal pipeline. The reasons to combine are express and field-specific: Cox directs its human-population training method to driverless automobiles; Ross's stated problem is precisely that pure behavior prediction ignores what a road user wants to do; Huval supplies the scale-collection infrastructure; Munro supplies the aggregation arithmetic. The result — a road-user predictor trained on aggregated human judgment — is the '687's stated objective and involves no change in system topology.
• Claims 2–10 are separately obvious, with claim 2 (the enumerated supervised-learner menu), claim 5 (Ross's action list and likelihoods), claim 6–7 (Cox's psychometric parameters), and claim 10 (the specification's own back-propagation error-signal loop) being the most cleanly met. Claims 3–4 (pixel manipulation; looped frames) are the weakest links in the chain on the art I verified and would need dedicated support.
• The '687's exposure is greater than the litigated family members' exposure, because the '687's independent claims omit the vehicle-control and state-of-mind limitations that the '344/'889/'046 add, while resting on the identical specification. A validity challenge that has been instituted against the narrower descendants has, if anything, a cleaner path against the broader parent.
• The dominant uncertainty is not the law; it is the dates and the Cox disclosure. Huval's, Zhang's, Ogale's and Munro's effective filing/publication dates are unverified and must be confirmed before any of this is relied upon; Zhang in particular is at risk of being post-dated to a July 5, 2017 critical date. And the entire ground structure for claim 1 turns on whether Cox fairly teaches aggregating human judgments of future road-user behavior into a training target, as distinct from aggregating object-label annotations.

Sources relied on: patents.google.com/patent/US10402687B2/en (specification, Background, Summary, FIG. 6/7/8 disclosures); justia.com patent 10402687 (claim listing); ptacts.uspto.gov PTAB petitions and Patent Owner Preliminary Responses for IPR2025-01573/-01574/-01575/-01576 (Tesla, Inc. v. Perceptive Automata LLC) — including Ground 1 (Ross + Cox), Ground 2 (Ross + Cox + Ogale), Ground 3 (Zhang), the Huval-Ross-Cox-Munro grounds, the printed-matter position, and the petition Fintiv/redundancy arguments; patents.justia.com/patent/10496091 and the front matter of US 11,709,490 (confirming Ross's Aug. 17, 2016 filing date via app. 15/239,367); courtlistener.com docket 72086670 (Perceptive Automata LLC v. Tesla, Inc., 7:25-cv-00594) for the secondary-considerations theory. All dates for Huval, Zhang, Ogale, Munro, Djuric, AA, Rolfe and Ellenbogen are unverified and are flagged as such in §2.