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.
1.1Surface measurements from a probe
What ifWhat if the input consists of contact probe coordinates and surface color readings?
Claim language“a plurality of images of an object”
A coordinate measuring arm records surface points and edge locations as a touch probe contacts each rigid part. A meshing routine fits surfaces to those coordinates and assigns colors from spot reflectance readings taken at registered locations.
1.2Joint positions from encoders
What ifWhat if the demonstration input is a time series of joint positions measured while a person operates a bench fixture?
Claim language“a video of a human manipulating the object”
The fixture clamps the object base and measures each hinge angle or slide displacement with an encoder at 100 samples per second. Forward kinematics combines each sample with calibrated joint axes to recover the position and orientation of each moving part.
1.3Geometry from fabrication files
What ifWhat if a fabrication file supplies the complete object geometry?
Claim language“reconstructing 3D object geometry”
The system loads a dimensional assembly file containing solid models, joint axes, and travel limits. It tessellates the solids for rendering and maps colors sampled from the object photographs onto the resulting mesh.
1.5Repeated motions recorded directly
What ifWhat if each object trajectory comes from a separate physical demonstration?
An operator repeats the task from different starting positions while calibrated cameras track fiducial markers fixed to the moving parts. The system stores the measured poses for each trial and pairs those poses with a robot motion planned for that trial.
1.7Training records from paired motions
What ifWhat if the paired motions produce numeric observation and action records for policy training?
Claim language“rendering videos”
A sampler evaluates the paired trajectories at 20 Hz and records robot joint positions, gripper opening, and object part poses as observations. It stores the planned joint and gripper commands as action targets, and the deployed policy receives the same observation fields from encoders and object pose tracking.
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
Robotic operation of recycling bin latches
A recycling station uses a robot to unlatch and open reusable sorting bins. Still images provide geometry and surface textures for the bin body, latch, and lid, while a video of a worker opening a bin provides the position and orientation history of each part. The generator places the recorded motion at different bin positions and computes robot joint trajectories that keep the gripper aligned with the latch and lid handle during their respective contact intervals. A renderer produces wrist camera clips, and a robot policy trains on the clips paired with their joint commands.
Component substitution
Direct keypoint prediction of part poses
A convolutional keypoint detector replaces the part motion extraction stage and predicts image coordinates for fixed landmarks on each moving part. A textured mesh reconstructed from the still images supplies the corresponding three-dimensional landmark coordinates and labeled renders for detector training. A perspective-n-point solver uses the detected landmarks and calibrated camera intrinsics to compute each part's position and orientation in each frame of the human demonstration. The resulting pose sequences drive object trajectory synthesis and robot trajectory generation for rendering and policy training.
Textured triangle meshes for object rendering
A textured triangle mesh represents each rigid object part in the renderer. Structure from motion estimates the still camera poses, and multiview stereo produces depth maps that are fused into a surface and triangulated. The pipeline projects color from the still images into a texture atlas and stores each moving part as a separate mesh in its own coordinate frame. During video generation, the renderer applies each sampled part transform to its mesh and draws the robot at the associated joint positions to produce policy training clips.
Parameter or scale variation
Fixed rate generation across fixture positions
A generator samples a four-second part pose history at 30 frames per second and translates the initial fixture position over a five-by-five grid with 20 millimeter spacing. At each grid point, it applies the same translation to every part pose and the associated gripper targets. A kinematic planner computes the robot joint sequence for each grid point and discards sequences that exceed joint limits or produce collisions outside designated contact surfaces. Each retained sequence yields a 120-frame clip paired with robot commands at the same sampling rate.
Combination with existing systems
Training pipeline linked to machine vision inspection
An industrial vision inspection station supplies calibrated camera poses, lens parameters, and part classes to a synthetic training pipeline for turning components before inspection. Still images provide the component geometry and texture, and a human demonstration video provides the component pose history during the turn. The pipeline generates robot and component trajectories for each part class and renders the executions from the inspection camera poses. The deployed robot policy takes the station's live images and selected part class as inputs and issues joint targets for presenting the required surface to the camera.
Automated or data-driven variant
Automatic contact timing from demonstration video
A hand landmark detector locates fingertips in each demonstration frame, and the tracked object poses locate the reconstructed part surfaces in the same camera view. Software identifies a contact interval when a fingertip stays within six pixels of a projected part surface and moves with the part for five consecutive frames. The trajectory generator attaches the gripper target to the moving part coordinate frame during each detected interval and interpolates between targets during the intervening frames. The renderer pairs video frames with the resulting gripper commands and contact intervals for policy training.
Specific version of a part
Turntable with indexed stops
A fixed camera photographs the object at 24 turntable stops spaced 15 degrees apart. The system repeats the sequence at a second camera height and stores the turntable angle and camera calibration with each image.
Stereo tracking of part markers
Two synchronized cameras record a rigid cluster of fiducial markers attached to each moving part. The system triangulates the visible marker centers and fits a rigid transform to the calibrated cluster, yielding three translation coordinates and a rotation for each frame.
Inverse kinematics with joint limits
The planner applies a fixed grasp transform to each object part pose to obtain the corresponding end-effector target. It solves inverse kinematics using the previous joint configuration as the seed, checks joint limits and collisions with the work surface, and adjusts the timing to satisfy joint velocity limits.
Video supervised action prediction
A visual encoder processes the four most recent rendered frames, and a regression head predicts the next joint velocity and gripper commands. Training minimizes squared error between the predictions and the time-aligned commands recorded from the robot trajectories used to render the videos.
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.
The result is that we can reconstruct an object model with curvilinear structures from a sparse set of images and can produce realistic renderings of the object model from arbitrary viewpoints.
A · US 2005/0140670 A1 · ¶ 45
The computer-implemented method of claim 1, wherein receiving training data includes receiving at least one of: an offline dataset that is associated with the plurality of atomic actions captured at particular time steps in an offline training environment and image data that is based on images by at least one camera that is associated with the plurality of atomic actions captured at particular time steps in the offline training environment.
B · US 2023/0316126 A1 · claim 2
1.2receiving a video of a human manipulating the object to perform a task;
1.3reconstructing 3D object geometry and an appearance of the object based on the plurality of images;
The curves and surface patches are further discretized to produce a triangle mesh for the object (FIG. 5(f)); Texture maps for the triangle mesh are generated from the original input images for synthetic rendering (FIGS. 5(g)&(h)).
A · US 2005/0140670 A1 · ¶ 53
1.4extracting 6 degree of freedom part motion of the object and its parts based on the video;
In one embodiment, data associated with the plurality of atomic actions may be captured in real-time by one or more cameras (not shown). The one or more cameras may be configured to complete vision based object tracking of the plurality of atomic actions that may be completed in order to complete a grasp sequence by the robot appendage 108 in an offline training environment (not shown).
B · US 2023/0316126 A1 · ¶ 35
1.5synthesizing multiple object trajectories based on the 6 degree of freedom part motion;
1.6generating associated robot trajectories for a robot to perform the task;
1.7rendering videos of a plurality of robot executions of the robot manipulating the object based on the multiple object trajectories and the associated robot trajectories; and
1.8training a robot policy based on the videos of the plurality of robot executions.
The one or more cameras may be configured to complete vision based object tracking of the plurality of atomic actions that may be completed in order to complete a grasp sequence by the robot appendage 108 in an offline training environment (not shown).
B · US 2023/0316126 A1 · ¶ 35
An empty cell means only that no passage was reproduced there.
Statements in these documents
Even with a descent tuning of these parameters, generally DRL algorithms may take a long amount of time to converge a policy.
B · US 2023/0316126 A1 · ¶ 2
According to one aspect, a computer-implemented method for providing accelerated reinforcement training.
B · US 2023/0316126 A1 · ¶ 3
Cite aspatent.fyi, US 2026/0355948 A1, claim 1, first published 2026-11-19T17:33:15-05:00, https://patent.fyi/US20260355948/, SHA-256 02dea1f4caae0a45b3c6eb725a1fd7bb6467eb12e2dacca3a12847034da9445b