The core classes are B25J for manipulators and robot arms, G06T for image data processing and generation, and G06V for image and video recognition, with G05B covering the control and regulating systems that link perception to motion. In this dataset B25J appears in 44 of 57 records, making it the dominant classification, while G06T and G06V together account for 24 records as the perception layer. Searches limited to a single class will miss the cross-disciplinary nature of most filings, since the majority combine a manipulator classification with at least one imaging classification.

The dataset shows filings rising from 2 in 2017 to a peak of 15 in 2019, then declining to 8 by the 2022 midpoint. This pattern is typical of a field where foundational claims around pose estimation, hand-eye calibration and grasp planning were staked out early, leaving later filers to refine rather than establish core methods. It is worth noting that publication lag of roughly 18 months means the most recent years in any such trend are always undercounted, so the apparent decline after 2019 should not be read as a definitive end to filing activity.

The tracked assignee list includes industrial automation names such as Siemens and Rockwell, specialist vision vendors such as Robovision, and academic filers including the Vakanski/Janabi-Sharifi pairing, which is the strongest co-assignee link in the dataset at five shared records. None of the tracked assignees show filings in the latest year, which reflects publication lag rather than a clean exit from the field. A representative recent filing from FANUC on learned visual pose estimation shows large manipulator OEMs are also active participants.

This FANUC filing, published 2025-08-21, claims a method of training a neural network on camera images captured from varying robot-arm positions and lighting conditions to infer a relative tool-center-point pose used in visual servoing control. It represents a learned, rather than purely geometric, approach to pose estimation. Anyone developing a similar supervised-learning pipeline for pose inference feeding a visual servoing loop should review its claim scope closely, since it sits squarely in the G06T/B25J overlap that this landscape identifies as the field's active perception layer.

Relative to the dense B25J manipulator core, sub-areas such as multi-camera pose fusion for occlusion handling, cluttered-scene grasp re-planning, and cycle-time optimization under variable lighting show thinner claim coverage in this dataset. These sit at the intersection of the smaller G06T, G06V and G05B classes rather than in the crowded manipulator-control layer. A first claim in these areas would likely need to specify a concrete technical mechanism, such as a fusion algorithm or re-planning trigger, rather than a generic vision-guided robot system, to clear the existing prior art.