patent.fyi
Applications  /  US 2026/0356394
United States Patent Application Publication

US 2026/0356394 A1

Published
Nov 12, 2026
Filed
Jul 21, 2026
Earliest date on the face
Dec 28, 2022
Classification
B60W 60/00272
First published here
Nov 20, 2026

Generated automatically. Not reviewed by a person. Not legal advice or an opinion.

Section 1 · machine-generated

What if?

Questions about how individual parts of claim 1 might be changed. A question may describe something different from what the claim language recites, or a version of it; no view is offered on which.

1.1Destination points with arrival windows

What ifWhat if each potential goal contained a destination point and an arrival time window?
Claim language“a goal path to the potential destination location”

A lane map supplies candidate destination points at lane exits and stop lines. The goal generator estimates an arrival window from the actor’s current speed and acceleration limits and stores each candidate as a point with earliest and latest arrival times.

1.2Raster convolution for scene encoding

What ifWhat if a raster convolutional network encoded the actors and their potential goals?
Claim language“in a graph neural network”

An overhead grid stores lane boundaries, actor positions, velocity components, and candidate path masks in separate channels. A convolutional encoder produces a spatial feature map, and pooling over actor footprints and candidate path masks yields actor and goal features.

1.2Temporal overlap at path crossings

What ifWhat if the interaction encoder represented path crossing conflicts with occupancy time bins?
Claim language“respective distances between a set of the plurality of actors and a shared space”

A motion estimator computes entry and exit times for each actor at the region where its candidate path crosses another path. The encoder marks occupied 100 millisecond bins over a five second horizon and combines the masks to identify simultaneous occupancy.

1.3Direct motion mode prediction

What ifWhat if the model predicted weighted motion curves directly from scene features?
Claim language“a forecasted goal”

A decoder converts each actor’s scene features into three weighted cubic splines constrained to start at the actor’s measured position and velocity. The vehicle planner checks each spline against its candidate motion over the forecast horizon.

1.3Discrete occupancy time slices

What ifWhat if the predictor returned goal-conditioned occupancy grids at fixed time steps?
Claim language“a continuous trajectory”

The predictor produces one grid of occupied-cell probabilities for each 200 millisecond step over a five second horizon. The planner checks the vehicle footprint against the grids and expands occupied cells using the actor’s speed bound to account for motion between steps.

Section 2 · machine-generated

Variations

Further arrangements of the same subject matter, described here as of the publication date at the foot of this page.

Other field of use

Forecasting at warehouse aisle crossings

An autonomous pallet carrier uses range sensors to track carts and workers and a floor map to locate aisle junctions and loading bays. It pairs each tracked actor with candidate destinations and routes, then marks shared crossings where pairs of routes intersect. A graph neural network encodes each actor's remaining route distance to those crossings and predicts a destination and continuous trajectory for each actor. The carrier controller uses the predicted destinations and trajectories to adjust its approach speed and commands motion when the crossing is clear.

Component substitution

Gradient boosted trees for goal selection

A gradient boosted decision tree ensemble replaces the graph neural network and receives one feature row per actor and candidate destination. Each row includes actor speed, heading error relative to the candidate path, and the two actors' distances along their respective paths to the nearest shared crossing, with a flag for paths that share no crossing. The ensemble scores each candidate and predicts speed and acceleration at fixed time knots for the selected path. A cubic spline interpolates distance traveled along the path, and the vehicle controller uses the selected destination and resulting continuous trajectory to set its approach speed.

Parameter or scale variation

Short horizon for depot traffic

A depot vehicle limits forecasting to 12 tracked actors within 30 meters and generates up to four destination and path pairs per actor. Every 100 milliseconds, it finds crossings among the candidate paths and supplies actor and goal nodes with the remaining path distances to each shared crossing. The model selects a goal and outputs a continuous trajectory over a three-second horizon. The motion controller checks those trajectories against its planned path and caps commanded speed at two meters per second.

Combination with existing systems

Forecasting with traffic signal timing

A roadside radio receiver supplies the forecasting computer with the current signal phase and the scheduled time of the next phase change for each controlled approach. The computer associates each candidate path with its signal group and adds phase and timing data to a graph that records each actor's distance along its path to a shared crossing. The model selects a goal and generates a continuous trajectory for each actor using the resulting graph features. The vehicle planner checks these trajectories for crossing conflicts and commands motion under the signal indication for its own approach.

Automated or data-driven variant

Learning candidate destinations from recorded tracks

An offline process aligns recorded actor tracks with a lane map and groups their endpoints by approach lane and exit lane. It stores the median endpoint and a representative path for each group in a table indexed by approach lane. During operation, the forecaster looks up candidate destination and path pairs for each tracked actor and encodes the distances remaining along the paths to shared crossings. The forecasting model selects a goal and generates a continuous trajectory, which the vehicle controller uses to time its entry into the crossing.

Generating paths from temporary lane markings

A camera segmentation model identifies temporary lane markings and traffic cones, and a range sensor locates them in vehicle coordinates. The forecasting computer fits corridor boundaries to successive detections and samples centerline paths from each actor's current position to reachable corridor exits. It locates crossing regions where candidate paths intersect and passes each actor's remaining distance to those regions into the forecasting graph. The model selects a goal and generates a continuous trajectory for each actor, and the vehicle controller uses the forecasts to command speed through the temporary corridor.

Specific version of a part

Lane graph destination sampling

A goal generator follows reachable lane segments up to 80 meters ahead of each tracked actor and places candidate destinations at lane ends and stop lines. Each candidate stores a destination coordinate and the concatenated lane centerline segments from the actor’s current position to that coordinate.

Graph features for crossing distances

A graph builder creates actor nodes and candidate goal nodes and computes a shared crossing polygon from overlapping candidate path corridors outside the actors’ current footprints. The graph network aggregates each actor’s path distance to the polygon into the actor and goal node embeddings.

Goal conditioned spline decoding

A scoring head ranks the candidate goals from their feature embeddings and selects the highest scoring goal. A trajectory head uses the selected goal embedding to generate a continuous piecewise cubic curve that starts at the actor’s measured position and velocity and ends at the selected destination.

Yield then controlled launch

The motion planner uses the actor’s forecasted goal to identify the crossing region and its continuous trajectory to estimate when the actor’s rear edge clears that region. The controller keeps the vehicle stopped until the predicted clearance time plus one second, then releases the brake and commands forward acceleration.

Section 3 · quoted from USPTO records

Other work

Passages reproduced word for word from two earlier-published U.S. patent documents, placed beside the language of claim 1. No characterization is offered. Each passage is marked A or B for the document it comes from; ¶ numbers are the paragraph numbers in the USPTO’s published full text of that document.

Document AUS 2021/0009163 A1
Published
Jan 14, 2021
Filed
Mar 12, 2020
Class
B60W 60/0027
Document BUS 2022/0153313 A1
Published
May 19, 2022
Filed
Nov 18, 2020
Class
B60W 60/00274
Claim 1
Document A
Document B
1.1determining a plurality of potential goals comprising at least one potential goal for a plurality of actors within an environment of an autonomous vehicle, wherein a potential goal comprises a potential destination location in the environment and a goal path to the potential destination location;
The method of claim 1 further comprising: by the presence detection subsystem of the autonomous vehicle, detecting a plurality of additional non-Ego actors; and by the on-board computing system of the autonomous vehicle at the point in time t0: predicting a plurality of possible modalities for each additional other actor to yield a combinatorial set of possible joint modalities of all non-Ego actors, reducing the set of possible modalities to yield a reduced set, predicting,
B · US 2022/0153313 A1 · claim 7
1.2processing the plurality of potential goals with a machine-learned forecasting model to generate a plurality of feature embeddings for respective nodes in a graph neural network corresponding to the plurality of actors and the plurality of potential goals, the plurality of feature embeddings indicative of respective distances between a set of the plurality of actors and a shared space distinct from spaces corresponding to the plurality of actors and corresponding to an intersection of goal paths associated with the set of the plurality of actors;
The node states can represent or describe the respective nodes' future trajectories based on their “perception” of the other actors. Such messages can be transposed into a frame of reference of the node receiving the message and/or can describe relative distances between the nodes.
A · US 2021/0009163 A1 · ¶ 27
1.3processing the plurality of feature embeddings to determine a forecasted goal for a respective actor of the plurality of actors and a continuous trajectory for the respective actor based on the forecasted goal; and
Respective trajectories of each actor can be predicted in a relational fashion with respect to each actor's nearby actors.
A · US 2021/0009163 A1 · ¶ 78
Returning to FIG. 2, at 205 Ego's prediction subsystem will again predict possible actions of Other, except in this step the predictions will be reactive forecasts to the possible actions of the virtual doppelganger, i.e., p(Other|Egodoppelganger=Ti,doppelganger) for each doppelganger trajectory Ti, doppelganger. The system will therefore forecast a set of trajectories F={f1, f2, . . . f3j} for each modality MOi.
B · US 2022/0153313 A1 · ¶ 32
1.4initiating a motion of the autonomous vehicle based on the forecasted goal and the continuous trajectory of the respective actor.
The means can be configured to determine a motion plan for the autonomous vehicle based at least in part on the motion forecast data.
A · US 2021/0009163 A1 · ¶ 41
A motion planning subsystem of the AV may then use the predictions to select actions that the AV will take in response to the moving actor's action, such as speed up, slow down, turn or move into another lane, etc.
B · US 2022/0153313 A1 · ¶ 20

An empty cell means only that no passage was reproduced there.

Statements in these documents
In particular, the present disclosure is directed to systems and methods for generating motion forecast data for actors with respect to an autonomous vehicle and training a machine learned model for the same
A · US 2021/0009163 A1 · ¶ 2
However, the interactions between various third party actors can be complex and difficult to model.
A · US 2021/0009163 A1 · ¶ 3
When an autonomous vehicle (AV) control system makes decisions about how and where to move the vehicle in an environment, a challenging problem is how to plan for possible interactions between the AV and other actors in the environment.
B · US 2022/0153313 A1 · ¶ 1
The space of possibilities grows exponentially when considering just one actor, and grows even faster when considering the actions of multiple actors; it quickly becomes computationally intractable to reason over the entire space of possibilities expeditiously.
B · US 2022/0153313 A1 · ¶ 3
Cite aspatent.fyi, US 2026/0356394 A1, claim 1, first published 2026-11-20T01:05:31-05:00, https://patent.fyi/US20260356394/, SHA-256 11161d22583cf4231c59012504606a90a840ee425de85e4e145c91342e7431ff
First published
Permanent addresshttps://patent.fyi/US20260356394/
Content fingerprint (SHA-256)11161d22583cf4231c59012504606a90a840ee425de85e4e145c91342e7431ff
Methodv2026-11-18.2 · automated, not reviewed by a person
Correctionscorrections@patent.fyi