This dataset returns 15 published patent families matching soft robotics, soft actuator and learning-based control terms across IPC classes B25J, G06N and B25J18. That is a small corpus compared with mature robotics fields, reflecting how recently reinforcement learning and sim-to-real methods have been applied to soft actuation. The true figure filed to date is likely somewhat higher, since publication typically lags filing by about 18 months and the most recent year is only partially represented.

The assignee base in this dataset is dominated by universities and research institutes rather than commercial companies, including groups such as the University of Alabama, Zhejiang University, and the Gwangju Institute of Science and Technology. No single assignee shows growth into the most recent tracked year, and co-assignee pairings are rare, appearing in only two of the fifteen families. This points to an academic, pre-commercialisation stage rather than one led by established robotics manufacturers.

General soft robotics patents cover actuator materials, mechanical design and manufacturing methods without necessarily involving a learning component. The subset tracked here requires both a soft robotics term and a learning-based control term such as reinforcement learning, sim-to-real transfer or morphological computation, narrowing the corpus to filings that combine physical soft actuation with an AI or machine-learning control layer. That combination is still uncommon: most records in this dataset classify primarily under manipulator hardware codes, with AI-model classification as a secondary rather than lead claim.

Several adjacent branches carry only one record each in this dataset, including printed circuit integration, harvesting equipment, surgical tooling and general control systems, all showing single-digit representation next to the sim-to-real and morphological computation search terms. That thinness suggests these are under-claimed rather than exhausted areas, particularly claims combining a specific soft actuator geometry with a named learning method such as reinforcement learning applied to a harvesting or surgical end-effector. A first mover with a well-scoped claim in one of these branches would face relatively little prior art within this search corpus.

US20230364777A1, assigned to the Board of Trustees of the University of Alabama, is the most-cited record in this dataset and claims a modular soft robot arm built from a base, an actuator, and a channelled arm structure with opposing surfaces. It does not block soft robotics broadly, but any design using a similarly channelled, modular arm geometry with a deforming actuator should be checked against its specific claim language before filing or launch. Because it is the clear citation leader here, it is the natural starting point for a freedom-to-operate review in this space, though a full clearance search should not stop at this one family.