This dataset identifies 72 patent families published between 2015 and mid-2026 that combine legged-robot terminology with terrain perception, 3D vision, obstacle detection or visual odometry claim language, filtered to specific IPC classes for image recognition and leg-drive mechanisms. That is a narrow, purpose-built slice rather than a count of all robotics or all computer-vision patents, so broader searches using different keyword combinations will return larger numbers. The 72-family count should be read as the intersection of legged locomotion and vision-specific claiming, not the total activity in either field alone.

Filing peaked at 14 in 2017 and has declined since, sitting at 2 by the 2022 midpoint and remaining low through the most recent year. This is a flat-to-declining trend rather than a growth curve, though the most recent year is always undercounted because publication lags filing by roughly 18 months. The pattern looks like an early cluster of foundational filings around 2017 that has not been followed by a comparable second wave, rather than a field that peaked and is actively winding down.

By citation count within this corpus, the leading records are mechanical patents on leg actuation and gait control, including a screw-actuator patent and a gait-behaviour patent, both cited well above other records in the set. Boston Dynamics also holds the newest highly-cited entry, a 2022 filing on detecting negative obstacles, which is the clearest vision-specific standout among the top-cited group. Citation counts favour older patents by nature of how citation accumulates over time, so these should be read as historically influential rather than necessarily the strongest current claims.

US20220390952A1, filed by Boston Dynamics and published in December 2022, claims a method for a legged robot to detect a candidate support surface below its current surface, check whether that lower surface has a gap in terrain data large enough to matter for a footstep, and if so mark that area as a no-step region. In practice this covers a specific technique for flagging holes, drops and similar negative obstacles so the robot avoids placing a foot there. Anyone building terrain-classification logic for legged robots should review this claim closely, since it is narrowly drawn around missing-data detection rather than obstacle detection generally.

The clearest gap is in dedicated vision claiming: G06V, the image-recognition IPC class, appears in only 5 of the 72 families, while the mechanical classes B62D and B25J appear in nearly all of them. That suggests vision functionality in this field is mostly claimed as a supporting feature within a locomotion patent rather than as a standalone method, leaving room for claims that isolate the vision or perception technique itself — for example fusing visual odometry with leg force sensing, or handling occlusion during obstacle detection — independent of a specific mechanical leg design.