Trajectory Optimization for Adaptive Informative Path Planning with Multimodal Sensing

TL;DR

This paper addresses adaptive informative path planning under resource constraints for an agent equipped with multimodal sensors, using a Gaussian process to model the unknown environment and a variance-reduction objective to guide exploration. It introduces a projection-based trajectory optimization method (GP-PTO) that linearizes dynamics, computes a descent direction via discrete LQR, and projects candidate trajectories back to feasibility while optimizing over sensor choices and motion. The approach is evaluated on a rover exploration benchmark, showing competitive performance with state-of-the-art methods and strong scalability for long-horizon planning, including variance reductions up to the levels reported in the abstract, and improved RMSE of the environment belief. The work contributes a formal AIPPMS formulation, the GP-PTO algorithm, and an open-source implementation, enabling robust multimodal sensing planning for planetary and remote sensing missions.

Abstract

We consider the problem of an autonomous agent equipped with multiple sensors, each with different sensing precision and energy costs. The agent's goal is to explore the environment and gather information subject to its resource constraints in unknown, partially observable environments. The challenge lies in reasoning about the effects of sensing and movement while respecting the agent's resource and dynamic constraints. We formulate the problem as a trajectory optimization problem and solve it using a projection-based trajectory optimization approach where the objective is to reduce the variance of the Gaussian process world belief. Our approach outperforms previous approaches in long horizon trajectories by achieving an overall variance reduction of up to 85% and reducing the root-mean square error in the environment belief by 50%. This approach was developed in support of rover path planning for the NASA VIPER Mission.

Figures (5)

• Figure 1: High-level overview of the projection-based trajectory optimization approach. The trajectory and sample locations are initialized. A descent direction is then computed to modify the trajectory in the direction that maximally reduces the objective function. This trajectory may violate the dynamics of the agent so the candidate trajectory is then projected back into feasibility. This process is repeated until convergence.
• Figure 2: Examples of trajectories produced from each of the six methods considered in this work. The true map is shown on the left. The top row shows the posterior mean and the bottom row shows the posterior variance of the Gaussian process world belief after the agent has reached the goal state. The agent starts in the bottom left corner. The pink line indicates the path the agent took and the green dots represent drill sites. These trajectories were generated with $\sigma_s = 1.0$ and $b = 60.0$.
• Figure 3: The trace of the Gaussian process covariance matrix is shown on the left and the RMSE of the beliefs with respect to the true map is shown on the right. For even comparison we evaluated $\text{Tr}(\Sigma)$ and the RMSE all using the same Gaussian process setup, even though the Ergodic and Random methods do not use the Gaussian process for decision making. Each subplot shows the average and standard deviation from 50 simulation runs. The budget was varied from $b=30$ to $100$ and the spectrometer noise was varied from $\sigma_s = 0.1$ to $1.0$.
• Figure 4: Drill placements along the trajectory from 50 simulations.
• Figure 5: Expected improvement results with respect to the true map. The average and standard deviation from 50 simulation runs are shown.