Bench-NPIN: Benchmarking Non-prehensile Interactive Navigation
Bench-NPIN: Benchmarking Non-prehensile Interactive Navigation
Ninghan Zhong, Steven Caro*, Avraiem Iskandar*, Megnath Ramesh*, and Stephen L. Smith
University of Waterloo
Abstract
Mobile robots are increasingly deployed in un structured environments where obstacles and objects are movable. Navigation in such environments is known as interactive navigation, where task completion requires not only avoiding obstacles but also strategic interactions with movable objects. Non-prehensile interactive navigation focuses on non-grasping interaction strategies, such as pushing, rather than relying on prehensile manipulation. Despite a growing body of research in this field, most solutions are evaluated using case-specific setups, limiting reproducibility and cross-comparison. In this paper, we present Bench-NPIN, the first comprehensive benchmark for non-prehensile interactive navigation. Bench-NPIN includes multiple components: 1) a comprehensive range of simulated environments for non-prehensile interactive navigation tasks, including navigating a maze with movable obstacles, autonomous ship navigation in icy waters, box delivery, and area clearing, each with varying levels of complexity; 2) a set of evaluation metrics that capture unique aspects of interactive navigation, such as efficiency, interaction effort, and partial task completion; and 3) demonstrations using Bench-NPIN to evaluate example implementations of established baselines across environments. Bench-NPIN is an open-source Python library with a modular design.
Environments
Navigating Maze with Movable Obstacles: This environment features a static maze structure with randomly initialized obstacles. The robot's task is to navigate from a starting position to a goal location while minimizing path length and obstacle collisions.
Autonomous Ship Navigation in Icy Waters: In this task, an autonomous surface vehicle must reach a horizontal goal line ahead while minimizing collisions with broken ice floes in the channel.
Box Delivery: The Box-Delivery environment consists of a set of movable boxes to be delivered to a designated receptacle. The robot is tasked to delivery all boxes using its front bumper.
Area Clearing: This environment consists of a set of movable boxes and a clearance area. The task of the robot is to remove all boxes from this clearance area.
Metrics
We propose a set of novel metrics to evaluate and compare policies that tackle navigation-centric and manipulation-centric NPIN. The proposed metrics capture properties unique to interactive navigation such as efficiency and interaction trade-offs, and partial task completions.
An illustration of three candidate paths and their performance. The shortest path (red) is the most efficient but suffers from excessive collisions. The longest path (blue) avoids collisions entirely but is inefficient. A balanced path (yellow) potentially offers the best trade-off.
Three candidate paths for area clearing, each achieving partial task success. Path 1 (red) is the most efficient but requires more effort due to unnecessary box movement. Path 3 (blue) minimizes effort by pushing as little as possible but results in a long path with low efficiency. Path 2 (yellow) offers a more balanced trade-off.
Reference Baselines
To support both general NPIN research and task-specific studies, Bench-NPIN includes a diverse set of baselines. These baselines fall into two categories: (i) reinforcement learning baselines applicable to all tasks and (ii) task-specific baselines for each task.
Reinforcement learning policies integrated with Stable Baselines 3
Planning-based policy using an ASV ice navigation lattice planner
Planning-based policy using a predictive ASV ice navigation planner
Spatial Action Maps policy
A Generalized Traveling Salesman Problem (GTSP) policy. Please see Appendix I for details