PhD Thesis Proposal
Carnegie Mellon University
Robotic systems are being leveraged to explore environments too hazardous for humans to enter. Robot sensing, compute, and kinodynamic (SCK) capabilities are inextricably tied to the size, weight, and power (SWaP) constraints of the vehicle. When designing a robot team for exploration, the diversity and types of robots used must be carefully considered because these variables have implications for ease of deployment. As the heterogeneity increases, so does the burden of tuning parameters to yield a performant team. To this end, a major challenge in robotic system design is to enable automatic adaptation of planning, mapping, and localization algorithms according to the SCK capabilities of the platform. Overcoming this challenge would significantly ease deployment of highly variable and diverse robots as a team. This proposal addresses the first two components of this challenge: designing adaptive planning and anytime mapping algorithms to enable scalable deployment of heterogeneous teams of robots for exploration.
To reduce the design parameters in the planning algorithms for exploration, a motion primitives-based action representation is employed due to its applicability to a diverse class of platforms. A key parameter in this paradigm is the maximum allowable speed along the motion primitives. This parameter is difficult to set for a heterogeneous robotic team because it is tied to the SCK capabilities. Existing methods assume the maximum speed to be fixed throughout exploration. However, in practice, this speed must also vary according to the complexity of the local obstacle distribution. To bridge this gap, a motion primitives-based adaptive-speed planning methodology is proposed that automatically sets the maximum allowable speed for the motion primitive library according to the changes in environmental complexity and robot SCK capabilities.
During the exploration task, a spatial map representation is incrementally created by streaming observations from a depth sensor. Metric and intensity information is often encoded into the map representation. The accuracy of this information affects the selection of the next most informative action. Moreover, the time taken to generate a high-accuracy map should typically be a fraction of the planning time. Therefore, an anytime mapping methodology is required that allows adaptive-resolution high-accuracy mapping depending on the SCK capabilities of the target platform. Most heterogeneous robotic teams use a discretized spatial representation of the environment. The key parameters in these techniques are the maximum spatial resolution of the map and specialized parameters depending on the optimizations employed to reduce the memory footprint. Just like the maximum speed parameter in planning, setting these mapping parameters is also difficult because it depends on the SCK capabilities. In this proposal, an anytime continuous mapping methodology is presented that enables adaptive-resolution occupancy inference and compute-aware high-fidelity spatial reconstruction from RGB-D point clouds within a time budget specified by the informative planner.
Finally, field deployment of an integrated exploration system with adaptive-speed planning and anytime mapping subsystems is proposed to quantify exploration performance gains using SWaP-constrained robots in complex cave environments.
Thesis Committee Members:
Wennie Tabib (co-Chair)
Nathan Michael (co-Chair)
Kostas Alexis, NTNU (Norway)