The objective of this project is to develop methods to design control systems for humanoid robots that show human levels of competence, robustness and flexibility in locomotion on human-scale rough terrain, focusing on responses to errors such as slipping and tripping, and responses to perturbations caused by irregular terrain. The research is motivated by the large disparity between human performance and current robot performance. The project uses a data-driven approach, utilizing motion data recorded from people to create trajectory libraries of these behaviors. It develops algorithms that allow these libraries to be adapted for robot control and generalized through interpolation, resequencing, and optimization for new environments. The project also explores strategy selection, modeling what strategies humans use in different situations. The project will have intellectual impacts in making better robots and understanding people better. The project will demonstrate better robot performance as well as more accurate models and simulations of human behavior. In both cases scientific publications will be augmented by extensive additional material and data on the web. Potential applications include insight into how the changes in the motor and sensory systems due to aging increase the risk of falling. The project may develop ways to change environments and train at-risk people to reduce the risk of falling.
This project originated as part of a long-term goal of developing methods to design control systems for humanoid, 2 legged, robots that mimic human behaviors during walking. Specifically, the project focused on recovery from balance disturbances, such as slipping and tripping, which would happen on irregular terrains. The research was motivated by the large disparity between human performance and current robot performance. People are much better at recovery from slips, trips and loss of balance than are current walking humanoid robots. The project utilized motion data recorded from people to create movement libraries of these behaviors in people. These movement responses are currently being analyzed for inclusion in robot walking through the development of algorithms allowing interpolation, resequencing, and optimization for new environments. This will result in balance recovery strategy selection, modeling what strategies humans use in different situations. Thus, the goal is to make better robots thorough understanding how people respond and also gain insight into how people control walking through the modeling effort. The focus of the work of the researchers at the University of Pittsburgh was to generate the human movement data in response to slipping and tripping and other conditions, and then develop a database of trajectories that can then be analyzed by our Carnegie Mellon University (CMU) partners towards implementation into new walking robot control. The major outcome of this research from the University of Pittsburgh was the data sets that were generated of the human movement recovery responses to slips and falls, along with subject demographics, balance evaluations and other subject-specific information. These datasets contain a collection of movement data that were collected in a number of human experiments. The movement data were broken down into trajectories of various parts of the body, such as legs, arms, head, and torso. For example, one database included responses to slips that occurred when stepping on a very slippery surface within the laboratory. Foot movements, arm recovery responses, and movements of the body were recorded and processed into trajectory libraries. Another dataset included responses to trips as people recovered. We also included context, such as whether the subjects knew the floor was slippery or was the condition of the floor unknown prior to the slip. Day-to-day changes in responses to repeated slips were also recorded. Similar context was supplied in the tripping database of responses. Subject demographic information was also included such as age (young, middle aged, and older adults), as well as the subjects’ functional balance, results of formal clinical balance testing, and range of motion. Other types of data included, when available, were electromyographic data (EMGs) from muscles in the lower legs that give an indication of when muscles are creating forces. Also, ground reaction forces generated under the feet walking from force plates were included. Finally, an additional data set was generated from further experiments done within the laboratory on patients with Parkinson’s disease and age-matched controls. This data will be useful in understanding responses when there are problems within the motor control system. These databases are now being analyzed by our CMU colleagues to evaluate the trajectories to build new approached to humanoid robot control. The techniques being used include machine learning and other trajectory-analysis methods. The results of this analysis will be finalized this coming year. We hope to also be able to then incorporate those models into our existing walking robots.
