Motion Planning for Heterogeneous Unmanned Systems under Partial Observation from UAV

Ci Chen, Yuanfang Wan, Baowei Li, Chen Wang, Guangming Xie, Huanyu Jiang

TL;DR

This paper introduces a novel motion planning approach for heterogeneous robot teams (UAVs and UGVs) operating with limited visibility. Their system, comprising a Grid Map Generation Network (GMGN) for perception and a Motion Command Generation Network (MCGN) with memory for decision-making, enables effective navigation in partially observed environments without needing to reconstruct a full map, achieving higher success rates than existing methods.

What Problem Does This Paper Address?

Coordinating unmanned aerial vehicles (UAVs) and unmanned ground vehicles (UGVs) offers significant advantages, with UAVs acting as aerial scouts to guide UGVs. However, a key challenge arises when the UAV's field of view (FOV) is insufficient to capture the entire environment, particularly in unfamiliar or disaster-stricken areas. This partial observability transforms motion planning into a Partially Observable Markov Decision Process (POMDP), a computationally complex problem. Existing solutions often require reconstructing a global map, which is time-consuming and may not be feasible in real-time. This paper aims to solve this by enabling effective motion planning for UAV-UGV systems under partial observation without global map reconstruction.

Methods

The proposed solution is a two-part system:

  • Grid Map Generation Network (GMGN): This network takes RGB images from the UAV's camera and generates a grid map representing the observed environment. It uses a lightweight ENet for semantic segmentation to identify pathways and obstacles, followed by a Post Processing Network (PPN) to convert these into a grid map format. This grid map captures the partial observation.
  • Motion Command Generation Network (MCGN): This is the core decision-making component. It leverages a combination of a Value Iteration Network (VIN) for planning and a Differential Neural Computer (DNC) with an external memory mechanism for reasoning and remembering important environmental features.
  • VIN: Inspired by traditional value iteration, this differentiable neural network learns to predict optimal actions by iteratively calculating state-action values.
  • DNC: This recurrent neural network with external memory allows the system to store and recall past observations and important landmarks, enabling it to reason about the environment and plan ahead, even with incomplete information.

The MCGN operates in two phases:

  1. Exploration Phase (before target is in FOV): Both UAV and UGV move synchronously. The MCGN uses the masked global observation matrix (derived from the partial observation and current position) and the DNC's memory to plan actions.
  2. Target Acquisition Phase (target in FOV): The UAV hovers, and the UGV moves to the target. The MCGN now uses the partial observation matrix directly, relying more on the VIN for planning.

Datasets

The paper does not explicitly mention specific public datasets used for training. Instead, it describes the process of creating labeled data for their simulation environment:

  • Semantic Segmentation Maps: Labeled using LabelMe software, distinguishing between obstacles (1) and passable parts (0).
  • Grid Maps: Also labeled as 0 or 1, derived from semantic segmentation maps.

The experiments were conducted in a custom simulation environment called HeROS, which is designed for heterogeneous unmanned systems.

Key Insights and Ideas

  • Partial Observation Planning without Reconstruction: The most significant insight is demonstrating that effective motion planning for heterogeneous systems is possible under partial observation without the need to build a complete global map. This significantly reduces computational load and improves real-time applicability.
  • Memory for Reasoning: The integration of an external memory mechanism (DNC) is crucial. It allows the system to "remember" important features and past observations, enabling it to reason about the environment and make more informed decisions, mimicking human exploration.
  • Two-Phase Planning Strategy: The distinction between an exploration phase (where both robots move) and a target acquisition phase (where the UAV hovers and the UGV moves) is a practical and effective way to handle the changing information landscape as the target becomes visible.
  • GMGN for Efficient Perception: The GMGN efficiently converts raw RGB images into a usable grid map representation, balancing accuracy with real-time performance.

Key Results

  • Improved Accuracy: The proposed method achieved an accuracy rate improvement of approximately 14% compared to the MACN baseline across various tasks with different obstacle configurations and observation ranges.
  • Higher Success Rate: The paper claims their method effectively plans motion and achieves a "relatively high success rate."
  • Reduced Collisions: Qualitative analysis showed that their method significantly reduced instances of collisions with obstacles compared to MACN.
  • Time Efficiency: In some scenarios, their method was able to reach the target point in fewer time steps than MACN.
  • Real-time Performance: The GMGN processed images in about 0.827s, and the MCGN issued commands in about 0.009s, meeting practical real-time requirements.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that in some scenarios (e.g., when the target is consistently outside the UAV's FOV and the UGV is forced into a repetitive downward movement), motion planning can fail because the system cannot transition to the target acquisition phase.
  • While the method generally performs well, one scenario showed that the planned path wasn't always optimal, involving unnecessary turns.
  • Future Work:
    • Further improvement of accuracy.
  • Extending the system to assist multiple UGVs with a single UAV.

Summary updated

Abstract

For heterogeneous unmanned systems composed of unmanned aerial vehicles (UAVs) and unmanned ground vehicles (UGVs), using UAVs serve as eyes to assist UGVs in motion planning is a promising research direction due to the UAVs' vast view scope. However, its limitations on flight altitude prevent the UAVs from observing the global map. Thus motion planning in the local map becomes a Partially Observable Markov Decision Process (POMDP) problem. This paper proposes a motion planning algorithm for heterogeneous unmanned systems under partial observation from UAV without reconstruction of global maps. Our algorithm consists of two parts designed for perception and decision-making, respectively. For the perception part, we propose the Grid Map Generation Network (GMGN), which is used to perceive scenes from UAV's perspective and classify the pathways and obstacles. For the decision-making part, we propose the Motion Command Generation Network (MCGN). Due to the addition of the memory mechanism, MCGN has planning and reasoning abilities under partial observation from UAVs. We evaluate our proposed algorithm by comparing it with baseline algorithms. The results show that our method effectively plans the motion of heterogeneous unmanned systems and achieves a relatively high success rate.
BibTeX
@inproceedings{chen2020motion,
  author = {Ci Chen and Yuanfang Wan and Baowei Li and Chen Wang and Guangming Xie and Huanyu Jiang},
  title = {{Motion Planning for Heterogeneous Unmanned Systems under Partial Observation from UAV}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2020},
  url = {https://openreview.net/forum?id=xvkRJ_VJThY},
}
Figure