Researchers Create Dataset of Human Activities for Robot Learning

TechXplore reported on Oct. 24, 2026, that researchers from Shanghai AI Laboratory, Zhejiang University and Shanghai Jiao Tong University had compiled a dataset covering 100 hours of human activities recorded with a wearable system. According to the report, the method was presented in an arXiv preprint.

The dataset compiled by the researchers was reported to combine video with body movements, finger movements, tactile information and descriptions of task procedures. The scope of recorded information is an issue in research that uses examples of human work to train robots, but the report does not establish whether this dataset improved learning performance.

The focus here is not robot performance or production, but a method for recording human tasks as data for robot learning. The reported dataset covers 100 hours and includes wearer-perspective video, reconstructed body movements, finger movements, available tactile information and descriptions of task procedures. It can be read as a design that handles multiple types of information related to movement and contact with objects, rather than relying on video alone. However, the report provided does not establish the recording accuracy of each type of information or how well they are aligned with one another. For tasks in which robots move while handling objects, a key focus is treating locomotion, posture and interaction with objects as parts of a continuous task. We covered IEEE Spectrum’s report on a quadruped robot autonomously solving a giant Rubik’s Cube, which addressed foot-end manipulation and whole-body control on the robot side. The research reported here concerns recording human movements and occupies a different stage: learning data rather than execution capability. No direct technical connection between the two has been established, but when considering robots’ work capabilities, how human movements used for training are recorded is a separate issue from body control. The paper described in the report is based on the idea that portable recording systems can be advantageous for collecting data in diverse environments. However, the total of 100 hours alone does not show the range of tasks, breadth of environments, number of participants or conditions for using the data. The article does not establish whether the dataset was actually used for training or improved performance compared with existing methods. The next points to examine are the recording conditions and annotation methods, access and usage terms, learning evaluations and validation on physical robots. See the original source for details on the technical setup and evaluation conditions.