Towards Safe Autonomy in Assistive Robots

Abstract

Robots have the potential to support older adults and persons with disabilities on a direct and personal level. For example, a wearable robot may help a person stand up from a chair, or a robotic manipulator may aid a person with meal preparation and housework. Assistive robots can autonomously make decisions about how best to support a person. However, this autonomy is potentially dangerous; robots can cause collisions or falls which may lead to serious injury. Therefore, guaranteeing that assistive robots operate safely is imperative.

This dissertation advances safe autonomy in assistive robots by developing a suite of tools for the tasks of perception, monitoring, manipulation and all prevention. Each tool provides a theoretical guarantee of its correct performance, adding a necessary layer of trust and protection when deploying assistive robots. The topic of interaction, or how a human responds to the decisions made by assistive robots, is left for future work.

Perception: Assistive robots must accurately perceive the 3D position of a person's body to avoid collisions and build predictive models of how a person moves. This dissertation formulates the problem of 3D pose estimation from multi-view 2D pose estimates as a sum-of-squares optimization problem. Sparsity is leveraged to efficiently solve the problem, which includes explicit constraints on the link lengths connecting any two joints. The method certifies the global optimality of its solutions over 99 percent of the time, and matches or exceeds state-of-the-art accuracy while requiring less computation time and no 3D training data.

Monitoring: Assistive robots may mitigate fall risk by monitoring changes to a person’s stability over time and predicting instabilities in real time. This dissertation presents Stability Basins which characterize stability during human motion, with a focus on sit-to-stand. An 11-person experiment was conducted in which subjects were pulled by motor-driven cables as they stood from a chair. Stability Basins correctly predicted instability (stepping or sitting) versus task success with over 90 percent accuracy across three distinct sit-to-stand strategies.

Manipulation: Robotic manipulators can support many common activities like feeding, dressing, and cleaning. This dissertation details ARMTD (Autonomous Reachability-based Manipulator Trajectory Design) for receding-horizon planning of collision-free manipulator trajectories. ARMTD composes reachable sets of the manipulator through workspace from low dimensional trajectories of each joint. ARMTD creates strict collision-avoidance constraints from these sets, which are enforced within an online trajectory optimization. The method is demonstrated for real-time planning in simulation and on hardware on a Fetch Mobile Manipulator robot, where it never causes a collision.

Fall Prevention: Wearable robots may prevent falls by quickly reacting when a user trips or slips. This dissertation presents TRIP-RTD (Trip Recovery in Prostheses via Reachability-based Trajectory Design), which extends the ARMTD framework to robotic prosthetic legs. TRIP-RTD uses predictions of a person’s response to a trip to plan recovery trajectories of a prosthetic leg. TRIP-RTD creates constraints for an online trajectory optimization which ensure the prosthetic foot is placed correctly across a range of plausible human responses. The approach is demonstrated in simulation using data of non-amputee subjects being tripped.