Belief Space Planning for Risk Signal Temporal Logic via Mixed-Integer Convex Programming

Key points

• In practice, robots estimate their state from noisy measurements, and when the information gained from a measurement depends on the state, the robot can actively reduce its uncertainty by planning in the belief space.1
• The resulting problem is, however, generally non-convex, for which gradient-based methods are neither complete nor globally optimal.1
• In this paper, we consider Risk STL (RiSTL) planning in Gaussian belief space and show that an information-form parameterization of the belief, together with a convex relaxation of the belief dynamics, yields a mixed-integer convex program (MICP).1
• We prove that the MICP is sound and complete, and that the relaxation preserves the globally optimal input trajectories.1

Sources (1)

• [1]Belief Space Planning for Risk Signal Temporal Logic via Mixed-Integer Convex Programming
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 12, 12:31 PM

  Signal Temporal Logic (STL) specifies robot tasks with spatio-temporal requirements, and its robustness semantics allow planning by trajectory optimization.

  In practice, robots estimate their state from noisy measurements, and when the information gained from a measurement depends on the state, the robot can actively reduce its uncertainty by planning in the belief space.

Extractive summary: sentences quoted from the sources.