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.
1.1Trajectories from rigid body simulation
What ifWhat if a simulator generates the interaction trajectories from a commanded motion and geometric models?
Claim language“accessing a recording”
A rigid body simulator advances gripper and object meshes in 2 ms steps using specified joint commands and surface friction. At each step, the controller sends the simulated gripper pose and solver contact coordinates directly to the trajectory encoder.
1.2Contact tracks on planar fixtures
What ifWhat if the system tracks fingertip contact points in a calibrated plane without reconstructing gripper poses?
Claim language“generating a sequence of three dimensional (3D) poses”
A guide constrains motion to a plane parallel to the face of a fixture-mounted part. An overhead camera tracks fingertip markers, and the controller maps their pixel coordinates to planar contact coordinates when pressure switches register contact.
1.2Net force and torque histories
What ifWhat if the system represents the interaction by its net force and torque without locating individual contacts?
Claim language“determining a corresponding set of contact points”
A six-axis load cell under a fixture measures the force and torque on the object at 500 Hz. The controller synchronizes those readings with the gripper poses and encodes the combined history for force-controlled replay.
1.2Direct comparison of contact trajectories
What ifWhat if the system compares stored contact trajectories directly without reducing them to embedding vectors?
Claim language“generating embeddings from the interaction trajectories”
A database stores each trajectory as timestamped contact coordinates in the object frame, grouped by finger. A dynamic time warping routine compares a query track with the stored tracks and retrieves the action with the lowest alignment cost.
1.2Embeddings computed for each request
What ifWhat if the controller computes each requested embedding from an archived trajectory and passes it directly to a decoder without maintaining a codebook?
Claim language“including the embeddings in a latent codebook”
An action identifier selects a timestamped trajectory from an archive. The encoder converts that trajectory into a temporary vector, and the decoder combines the vector with the target object's pose to generate a motion sequence.
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
Seedling transfer between nursery trays
A fixed stereo camera pair records a compliant two-finger gripper lifting seedling plugs from propagation cells and placing them into transplant trays. Visible fiducials on the gripper and tray register the reconstructed gripper poses to the cell grid, and the reconstructed jaw surfaces and plug boundary define contact points at each sample. The encoder stores the approach, squeeze, lift and release trajectories as codebook entries in a coordinate frame centered on the plug.
Component substitution
Cosine coefficients for trajectory encoding
A discrete cosine transform replaces the learned trajectory encoder. The processor resamples each demonstration to 64 samples and expresses jaw position, orientation and contact coordinates on each jaw in an object-fixed frame. It transforms each coordinate channel, retains the first eight coefficients and appends the start and end times of each continuous contact interval to form a codebook entry.
Parameter or scale variation
High speed recording of snap closures
Two synchronized cameras record gripper demonstrations at 1,000 frames per second with 100-microsecond exposures across a 40-millimeter field of view. The processor uses calibrated stereo views to reconstruct jaw poses and identify contact points from the jaw and latch surfaces. It interpolates the trajectories onto 128 samples spanning 32 milliseconds before and 95 milliseconds after the first latch contact, then encodes each window into the codebook.
Combination with existing systems
Codebook control of connector insertion
A connector assembly station pairs a camera-guided robot with a wrist force and torque sensor. An image locates the socket, and the controller transforms the selected insertion trajectory from socket coordinates into robot coordinates. During insertion, the controller uses wrist measurements to adjust lateral position and angle, pauses motion when axial force exceeds a configured limit and records the measured force alongside the replayed trajectory.
Automated or data-driven variant
Contact labels from tactile measurements
Each gripper jaw carries a calibrated tactile array synchronized to the demonstration video by a shared clock. Software groups adjacent sensor cells whose readings exceed a baseline threshold and calculates a pressure-weighted centroid for each group. The reconstructed jaw pose maps each centroid into the object coordinate frame, and the encoder combines centroid paths, pressure histories and gripper poses into codebook entries.
Automatic selection of representative demonstrations
Software groups recorded demonstrations by task label and aligns their pose and contact trajectories using dynamic time warping. The alignment cost combines object-relative jaw displacement, angular displacement and differences in contact state. A clustering algorithm selects the recorded trajectory with the lowest total alignment cost to the other members of each cluster, and the codebook stores that trajectory's embedding and the cluster's observed duration range.
Specific version of a part
Synchronized global shutter recording
Two global shutter cameras view the gripping structure and object from different angles at 120 frames per second. A common hardware trigger exposes both sensors at each frame interval, and the recorder stores both image streams with shared frame indices.
Marker based gripper pose fitting
A stereo camera pair triangulates three noncollinear markers on the gripper palm and additional markers on the moving fingers. A least-squares fit to the known gripper geometry estimates the palm position, palm orientation, and finger joint angles for each frame.
Mesh based contact estimation
For each fitted pose, the processor transforms sampled fingertip surfaces into the coordinate frame of a registered object mesh. It records the closest points on the object mesh for fingertip samples within a 1 mm distance threshold.
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.
1.1accessing a recording in which a gripping structure is performing an action against an object, wherein the recording includes time-series data illustrating performance of the action by the gripping structure over a period of time;
Various embodiments may receive and use sensor signals from various sensors, such as video, radar, LiDAR, ultrasound, movement, acceleration, heat map, etc., for example in order to acquire sensor data for detecting objects (i.e., other agents) and recording their previous trajectories and as input for the ML model for predicting the behavior.
A · US 2023/0169313 A1 · ¶ 90
1.2from the recording, generating interaction trajectories of the gripping structure performing the action by: generating a sequence of three dimensional (3D) poses of the gripping structure while the gripping structure is performing the action; and for each 3D pose included in the sequence of 3D poses, determining a corresponding set of contact points between the gripping structure and the object during performance of the action; generating embeddings from the interaction trajectories; and including the embeddings in a latent codebook.
At block 106, the controller 34 extracts the features of each trajectory 60 determined at block 104. Further, the controller 34 generates a trajectory feature vector for each trajectory 60.
B · US 2024/0140475 A1 · ¶ 30
An empty cell means only that no passage was reproduced there.
Statements in these documents
Accordingly, reliable approaches to predict agent behavior, i.e., to determine (expected) trajectories in a multi-agent scenario, are desirable.
A · US 2023/0169313 A1 · ¶ 5
The method described above allows for effective prediction of trajectories in a multi-agent scenario.
A · US 2023/0169313 A1 · ¶ 8
It is therefore desirable to develop a system for maintaining an autonomous vehicle within its current lane even when the lane markings and the road edges are not visible.
B · US 2024/0140475 A1 · ¶ 3
The method described in this paragraph improves autonomous vehicle technology by allowing a vehicle to autonomously maintain itself within a road lane even when the road edges and/or lane markings are not visible.
B · US 2024/0140475 A1 · ¶ 4
Cite aspatent.fyi, US 2026/0355960 A1, claim 1, first published 2026-11-19T17:33:15-05:00, https://patent.fyi/US20260355960/, SHA-256 8814af482a41fa72c173a50bc2c406b23263e6c86d8b9e428b169332cc533319