DARE: Diffusion Policy for Autonomous Robot Exploration

TL;DR

This paper proposes DARE, a novel generative approach that leverages diffusion models trained on expert demonstrations, which can explicitly generate an exploration path through one-time inference and builds DARE upon an attention-based encoder and a diffusion model.

Abstract

Autonomous robot exploration requires a robot to efficiently explore and map unknown environments. Compared to conventional methods that can only optimize paths based on the current robot belief, learning-based methods show the potential to achieve improved performance by drawing on past experiences to reason about unknown areas. In this paper, we propose DARE, a novel generative approach that leverages diffusion models trained on expert demonstrations, which can explicitly generate an exploration path through one-time inference. We build DARE upon an attention-based encoder and a diffusion policy model, and introduce ground truth optimal demonstrations for training to learn better patterns for exploration. The trained planner can reason about the partial belief to recognize the potential structure in unknown areas and consider these areas during path planning. Our experiments demonstrate that DARE achieves on-par performance with both conventional and learning-based state-of-the-art exploration planners, as well as good generalizability in both simulations and real-life scenarios.

Figures (6)

• Figure 1: An example planned path from DARE. Based on the robot belief (represented by the occupancy grid map), a robot (represented by the axes) constructed an informative graph. The graph is passed to an attention encoder network and a diffusion policy network to output a planned exploration path (in orange). The robot executes the planned path in the receding horizon manner until it fully explores the environment. Note DARE can reason about the partial belief to recognize the potential structure in some unknown areas and consider these areas during path planning.
• Figure 2: Diffusion-based exploration planner. At each step, DARE maintains a graph-based belief representation and encodes it with self-attention layers to capture a robot belief feature. Conditioned on a sequence of robot belief features, the diffusion policy generates future action sequences through iterative denoising. Note that planned paths can extend to unknown areas.
• Figure 3: DARE exhibits the ability to predict unknown areas. Here we show some examples where the predictions are correct. The previous trajectory of the robot (represented by the red dot) is in red. The planned path by DARE is in green. The explored free areas are in white while the unexplored free areas are in light gray.
• Figure 4: Trajectory analysis in the Gazebo simulation.
• Figure 5: Exploration paths comparisons in the Gazebo simulation.