patent.fyi
Applications  /  US 2026/0355967
United States Patent Application Publication

US 2026/0355967 A1

Published
Nov 12, 2026
Filed
Jun 25, 2026
Earliest date on the face
Dec 27, 2023
Classification
B25J 9/1687
First published here
Nov 19, 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.

1.1Camera only pose measurement

What ifWhat if the controller used only images to locate the socket and the held part?
Claim language“obtaining multi-modal data”

A fixed camera images fiducials on the socket fixture and on the gripper. The controller estimates the relative pose from their image coordinates and commands an approach along the socket axis. A spring-centered gripper mount takes up small lateral offsets during insertion.

1.1Socket location from live images

What ifWhat if the task specified the mating feature type and the controller found its position at run time?
Claim language“an assembly position indicated by an assembly task”

A depth camera locates the single empty socket by fitting a circle to its rim and a plane to the surrounding face. The fitted center and plane normal define the insertion path, while a wrist load cell measures axial force during seating.

1.2Separate alignment and seating controllers

What ifWhat if vision handled alignment and force handled seating in separate control stages?
Claim language“fusing the multi-modal data”

The vision controller sends a pose command to bring the held part above the socket. When the approach move completes, the insertion controller advances along the tool axis using wrist force as its sole sensor input. The insertion controller stops at a preset axial load, and the two controllers exchange only a stage completion signal.

1.3Rule based insertion sequence

What ifWhat if fixed transition rules selected the approach, search, and seating motions?
Claim language“a trained reinforcement learning model”

A state machine reads a fused estimate of lateral offset, insertion depth, and axial load. It commands a straight approach, switches to a planar spiral if axial load rises before insertion begins, and resumes axial feed when depth increases. It stops at the target depth or the configured force limit.

1.3Model predictive insertion control

What ifWhat if a constrained optimizer computed each insertion command from an explicit contact model?
Claim language“a trained reinforcement learning model”

The controller combines camera pose, encoder position, and wrist wrench into a relative pose and contact state. At each update, it uses a linear spring model of contact to predict the effect of candidate tool velocities over the next ten steps. It applies the first velocity from the solution that reduces pose error within the force and speed limits.

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

Snap fitting in footwear assembly

A robotic gripper holds a molded heel insert over a shoe sole clamped in a fixture. A depth camera measures the recess and insert poses, and a wrist force and torque sensor measures loads during contact. The controller fuses the relative pose and contact loads in fixture coordinates, and a trained reinforcement learning policy commands incremental gripper motions to engage the snap tabs. A displacement probe checks the seated height before the fixture releases the sole.

Component substitution

Model predictive control for pin insertion

A model predictive controller replaces the trained reinforcement learning model in a robotic pin insertion station. An overhead camera estimates the bore center and pin axis, and a wrist force and torque sensor measures contact loads. The controller fuses the pose estimates and loads in fixture coordinates and uses a local model of pin motion and contact to predict the effects of candidate tool movements. At each update it applies the first movement from the sequence that minimizes predicted alignment error and insertion depth error while respecting force and motion limits.

Parameter or scale variation

Low force insertion of miniature pins

A three axis positioning stage seats a 0.8 mm diameter pin in a bore with 40 micrometers of diametral clearance. A telecentric camera samples at 100 frames per second, and a load cell and stage encoders sample at 1 kHz. The fusion routine combines the latest image offset with the mean load and encoder displacement over each 10 ms interval. A trained reinforcement learning policy commands stage increments of up to 10 micrometers per update, and an independent load interlock stops the stage at 2 N.

Combination with existing systems

Connector insertion on an indexed conveyor

An indexed pallet conveyor brings housings to a robot station, where a programmable logic controller stops the pallet and closes a locating clamp. A station camera measures each connector opening relative to the robot gripper, and a wrist sensor measures insertion force and torque. The robot controller fuses the relative pose and contact loads and supplies them to a trained reinforcement learning policy that commands connector alignment and insertion. The conveyor controller releases the pallet after the robot reports the required seating depth and the inspection circuit confirms electrical continuity.

Automated or data-driven variant

Training with randomized fixture and contact parameters

A rigid body simulator represents the robot, gripper, part, and assembly fixture and generates camera frames, wrist loads, and joint positions. Training software samples fixture offsets, contact friction, and sensor noise at the start of each simulated episode, then fuses the simulated measurements in the fixture frame. A reinforcement learning algorithm trains a policy from rewards for seating the part and penalties for excessive contact loads. The physical station applies the same fusion routine to live measurements and uses the trained policy to command tool motions.

Automatic sensor weighting during insertion

A wrist camera tracks the part and opening while a wrist force and torque sensor measures contact and joint encoders report tool motion. A state estimator propagates the tool pose from encoder readings and adjusts the weight of each camera update using its feature reprojection error. The estimator supplies relative pose, pose uncertainty, and fixture referenced contact loads to a trained reinforcement learning policy. The policy commands small alignment and insertion movements as the gripper enters the opening and the visible area decreases.

Specific version of a part

Fixed camera and wrist sensing

The task supplies a nominal socket pose in the robot base frame. A fixed global-shutter camera captures the socket rim and gripper fiducials at 30 frames per second. A wrist transducer records three force and three torque components at 1 kHz, and the controller timestamps both streams against a common clock.

Time aligned observation vector

The fusion stage uses each camera exposure timestamp to interpolate joint encoder positions and select a 20 ms window of wrist force samples centered on the exposure. It transforms the estimated relative pose and the mean wrench into the tool frame, then concatenates them with joint velocities in one observation vector.

Contact aware pose filtering

An extended Kalman filter propagates the tool-to-socket pose from joint encoder increments and corrects it with camera pose observations. When the wrist force indicates contact, the filter uses the measured normal force and a calibrated contact stiffness to update its estimate of axial deflection. It outputs the relative pose, axial deflection, and their uncertainty estimates.

Learned incremental tool motion

A policy trained by reinforcement learning reads a fused vector of pose error, tool velocity, and wrist wrench at 50 Hz. It outputs incremental translations and rotations, with per-update limits of 0.5 mm per translation axis and 0.25 degrees per rotation axis. A joint position servo tracks the resulting tool pose during alignment and insertion.

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 11,623,339 B1
Published
Apr 11, 2023
Filed
Sep 3, 2019
Class
B25J 9/0009
Document BUS 2023/0154170 A1
Published
May 18, 2023
Filed
Nov 15, 2022
Class
G06V 10/806
Claim 1
Document A
Document B
1.1obtaining multi-modal data, wherein the multi-modal data comprises at least two of image modal data, force data, and motion data, and the multi-modal data comprises an assembly position indicated by an assembly task and information about a mechanical device that is to execute the assembly task;
The portable robotic manipulation assembly 500 may therefore be a robotic picking system that includes the first robotic arm 550 configured to grasp objects, a moveable camera assembly 560, one or more torque or force sensors, and a controller configured to control operation of the first robotic arm 550 and moveable camera assembly 560.
A · US 11,623,339 B1 · ¶ 63
1.2fusing the multi-modal data to obtain fused data; and
The method includes generating three-dimensional (3D) feature information and two-dimensional (2D) feature information based on a color image and a depth image, generating fused feature information by fusing the 3D feature information and the 2D feature information based on an attention mechanism, and generating predicted image information by performing image processing based on the fused feature information.
B · US 2023/0154170 A1 · abstract
1.3controlling, based on the fused data and by using a trained reinforcement learning model, the mechanical device to execute the assembly task.
The controller may be configured to: determine that the item sorting selectable option was selected, cause the boom to position the camera assembly to image a group of items, determine a first item of the group of items, and cause the first robotic arm to grasp the first item and move the first item from a first location to a second location.
A · US 11,623,339 B1 · ¶ 50

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

Statements in these documents
However, certain robotic arms may be in fixed locations or otherwise difficult to move.
A · US 11,623,339 B1 · ¶ 2
For example, due to power constraints, robotic arms may not be moveable and may therefore be designated to perform a specific task.
A · US 11,623,339 B1 · ¶ 2
If information of a single modality is used, the accuracy of the image processing result may be reduced.
B · US 2023/0154170 A1 · ¶ 3
Cite aspatent.fyi, US 2026/0355967 A1, claim 1, first published 2026-11-19T17:33:15-05:00, https://patent.fyi/US20260355967/, SHA-256 3d2908eff5105aa4e047261415a63afe5464c81b1dc5a670c84ccbce844ef090
First published
Permanent addresshttps://patent.fyi/US20260355967/
Content fingerprint (SHA-256)3d2908eff5105aa4e047261415a63afe5464c81b1dc5a670c84ccbce844ef090
Methodv2026-11-18.2 · automated, not reviewed by a person
Correctionscorrections@patent.fyi