Quantitative Path-Planning from Qualitative Language Instructions

Humans and robots think in different ways: humans evaluate qualitatively while robots process quantitative data. For example, a human may say “I am at the east of the building” in expressing a location, while a robot may use a precise coordinate. Similarly a human may specify “run carefully” in requiring a behavior pattern, but a robot may read linear and angular velocities. There is a need to convert between qualitative and quantitative representations for mutual understanding in human-robot interaction. Conversion from machine data to human-readable form may be possible through visualization and other tools, but it is still somewhat difficult for a robot to understand a human’s expression particularly where there is ambiguity. Understanding qualitative instructions may enable robots to work with untrained human users. We are interested in a robot that plans a path using information specified by a human using natural language; the path may contain multiple criteria and topological requirements. Planning paths from language instructions depends on • understanding the qualitative requirement given in a language instruction; and • being able to plan paths using preferences over multiple criteria and topological constraints. Our approach supports a robot to understand human intent in path-planning, which includes (a) a probabilistic model that converts an instruction into a path-planning problem, (b) path-planning algorithms that support multiple objectives and multi-topology constraints. Firstly, a language instruction is grounded into objectives and topology constraints in Section 2. Multiple criteria are supported by multi-objective pathplanning in Section 3, and topology constraints are supported by homotopy-based optimal path-planning in Section 3.