Ergonomically intelligent physical human-robot interaction
Ergonomically intelligent Teleoperation
Ergonomics and human comfort are essential concerns in physical human-robot interaction (HRI) applications. Common practical methods for postural estimation, assessment, and correction either fail in estimating the correct posture due to occlusion or suffer from incomplete ergonomics models. We are focusing on developing new artificial intelligence techniques to reduce the risk of musculoskeletal injuries in physical HRI that can overcome the above issues.
We introduced ergonomically intelligent physical HRI and teleoperation, which includes novel methods for
Posture estimation
Ergonomics risk assessment
Postural optimization
Postural correction
solely using the interacting robotic input device. Our intelligent and low-cost system benefits from artificial intelligence, machine learning, probabilistic modeling, and planning techniques.
In postural estimation, our approach uses only the trajectory of the interacting robot to estimate the human's posture with adequate accuracy compared to motion capture. In addition, we introduced DULA, a differentiable and continuous ergonomics model, learned by a neural network. The DULA model expands the utility of risk assessment methods such as the RULA and enables its use in gradient-based postural optimization for calculating the optimal postural correction in different types of physical human-robot interaction and teleoperation tasks. Finally, we introduced a new autonomous approach for ergonomics intervention and postural correction in teleoperation using the leader robot while the user continues performing the task.
Current posture estimation approaches such as motion captures and markerless methods are prone to occlusion in p-HRI and teleoperation, due to the proximity of the human and the robot, and require expensice devices and time-consuming setup.
Here, we propose a low-cost and probabilistic solution that can estiamte human posture (segments lengths and joint angles) solely from the trajectory of the interacting robot (leader robot in teleoperation). We show that our approach is accurate enough for continuous monitoring of the posture and risk of musculoskeletal disorders and can raise an alarm for further investigations if the risk of injuries goes high.
A computational model for human comfort and ergonomics is the key part in posture optimization reserach. Current literature either use quadratic approximations of common risk assessment tools or use some developed analytical models. However, the reliability and optimality of the level of ergonomic benefit they provide are questionable comparing to standard risk assessment tools in ergonomics. In addition, risk assessment tools such as RULA and REBA provide discrete scores, they are non-differentiable, and they include plateus which create challenges when using them in gradient-based postural optimization.
To solve the above issues, we propose DULA (differentiable upper limb assessment), a continous and differentiable model for ergonomics that is learned from a dataset of postures labeled by RULA using a neural network. It predicts the risk score with 90.73 % accuracy compared to RULA. We are also developing DEBA (differentiable entire body assessment) similae to DULA for full body assessment.
Neural network structure for learning DULA
Autonomous and intelligent postural optimization in p-HRI and teleoperation is a fresh trend in the robotics research. Researchers have tried different type of gradient-based and gradient-free optimization approaches to find the optimal postural correction of the human interacting with a robot in some sparse set of tasks.
In our reserach, we had a fundamental view over the problem. We categorized different types of p-HRI and teleoperation tasks and proposed the first general framework for postural optimization in p-HRI and teleoperation.
We utilized DULA model from our other work in gradient-based postural optimization and showed that how it can improve the posture and reduce the risk of injuries in a simulation-based study.
Postural optimization framework for teleoperation
To initial evaluation of our postural optimization approach, we developed a simulator that includes the human and the robot (two robots for teleoperation). The simulated human behaves like a human in two ways:
physically controlling the interactive task
accepting or rejecting the recommended optimal postural corrections
We model this as an optimal motion planning framework with re-planning that finds a human joint trajectory that interacts with the robot to do the task while moving toward the optimal ergonomic posture.
Overview of the simulated teleoperation
This project is suppprted by NIOSH ( award number T420H008414-10) and DARPA (under grant N6600119-2-4035).