UC San Diego

As robots become more integrated into society, their reasoning and actions willinvariably be evaluated by human decision makers. Thus, robots need to perceive, act, and reason like humans to maintain clarity, trust, and safety. In this thesis, we consider navigation problems, which consist of designing global planning and reactive control modules for single and multiple robots. While most current navigation strategies are robot-centric, here we take a human-centric approach and design navigation algorithms that are intuitive, risk-perception-aware, and possibly non-rational (as humans often are in risky situations). First, we focus on intuition and consider a formation control problem for a distributed robotic swarm. We develop a novel Human-Swarm Interaction (HSI) framework using the notion of an interpreter, enabling the user to control a robotic swarm’s shape and formation with intuitive hand gestures. The interpreter acts an intermediary, translating a high-level shape inputs to swarm specifications and vice versa. These specifications are then translated into commands, which are calculated and executed in a decentralized manner to depict the intended shape. Next, we focus on a single robot deployed in environments that contain generic moving sources of risk (for example, human-like obstacles requiring certain social distancing). We develop planning (via RRT*) and control (via CBFs) algorithms, that take human-like non-rational risk perception of the environment into account. We use Cumulative Prospect Theory (CPT), a non-rational model from Behavioral Economics, to construct perceived risks in the environment, capable of depicting a wide spectra of risk profiles. We introduce three new metrics: “Expressiveness”, “Inclusiveness,” and “Versatility” to characterize the richness of a risk model. We prove that CPT is superior in all these categories when compared to other popular models such as Conditional Value at Risk (CVaR) and Expected Risk (ER). This is further confirmed via simulations, which show that our approach can capture a richer set of meaningful paths, representative of different risk perceptions in an environment. We also observe that a learning algorithm using CPT can approximate the risk profile of arbitrary paths in an environment better than CVaR and ER. From a controls perspective, we prove that our CBF based approach result into larger feasible control set for a robot when using CPT. Finally, we propose a novel user study design to understand human path planning in everyday risky and uncertain environments. Considering a COVID-19 pandemic grocery shopping scenario, we ask participants to choose paths with varying risks (proximity to sick people) and time-urgency (path length). We reveal that participants in general are willing to take more risks and time-urgent paths, contrary to the popular assumption that humans are in general risk averse. Data analysis further shows that human decision making is better captured by CPT, as compared to CVaR and ER, thus validating our CPT approach to model non-rational risk perception in navigation problems.