Active Object Perceiver: Recognition-Guided Policy Learning for Object Searching on Mobile Robots

Xin Ye, Zhe Lin, Haoxiang Li, Shibin Zheng, Yezhou Yang

Introduction

In recent years, researchers have made great strides in "target-driven visual navigation," where a robot is given a picture of a scene and must navigate to the location where that picture was taken. However, asking a robot to find a specific object (e.g., "find a mug") based only on a generic reference image of that object is a much harder problem.

The target object might be tiny in the robot's field of view, viewed from an arbitrary angle, or hidden in a complex environment. Furthermore, the robot in this scenario does not have access to a global map or odometry; it must rely entirely on the live RGB video stream from its camera to figure out how to move and where to look.

Robot trajectory searching for a target object in a real-world environment.

Methodology

The authors propose a framework consisting of two main deep learning modules:

1. Object Recognition Module

Instead of a standard object detector, the authors use a two-stream ResNet-50 architecture. It takes two inputs: the robot's current camera view and an image of the target object. It outputs a bounding box if the target object is detected in the current view.

2. Recognition-Guided Action Policy (Deep RL)

This module decides the robot's next move (e.g., move forward, turn, tilt camera). It takes the current view, the target image, and a spatial mask of the bounding box (from the first module) as inputs.

Architecture of the recognition-guided action policy network showing the fusion of visual features and object location masks.

The Novelty: A Decaying Reward Function

Standard RL reward functions (e.g., giving a +10 reward at the goal and a -0.01 penalty for every other step) fail in this complex task because the robot rarely stumbles upon the goal by chance. To fix this, the authors designed a custom reward function based on the area of the detected bounding box.

Crucially, to prevent the robot from oscillating back and forth between two spots to continuously harvest rewards from a distant bounding box, the system only grants a reward if the current bounding box is larger than any bounding box seen previously in that episode. This forces the robot to actively approach the object.

Datasets and Evaluation

  • AI2-THOR (Augmented): The authors used the existing AI2-THOR 3D simulation environment. Because the original dataset lacked object-level bounding boxes for this specific task, the authors manually annotated ground-truth bounding boxes for target objects across 4 scenes.
  • Real-World Turtlebot Dataset: The authors collected a new dataset using a physical Turtlebot in a conference room. The room was discretized into 27 locations, with 8 pan/tilt RGB images captured per location, representing the physical state space of the robot.

Key Results

  • Convergence: Models trained with standard reward functions completely failed to converge, performing worse than a baseline random walk. The authors' novel decaying reward function successfully allowed the model to learn the task.
  • Simulation Performance: In a scenario searching for 4 different objects in a simulated scene, the high-exploration model achieved a 92.5% success rate (averaging 1643 steps). The low-exploration model averaged only 30.1 steps but dropped to a 75% success rate.
  • Real-World Efficacy: When deployed on the physical Turtlebot, the model achieved a 100% success rate. The high-exploration variant took an average of 176.8 steps, while the low-exploration variant took just 63.3 steps.

Discussion

The exploration vs. exploitation trade-off is highly visible in this task. High exploration leads to higher success rates at the cost of longer paths, while low exploration yields efficient paths but risks getting stuck in local optima.

The authors acknowledge two primary limitations for future research:

  1. Lack of Memory: The current system is purely reactive to its current frame. It lacks an explicit memory unit (like an LSTM or spatial map) to remember which parts of the room it has already searched.
  2. Lack of Common Sense: Humans use semantic priors to find things (e.g., "a plate is likely on a table"). Integrating structured common-sense knowledge could drastically speed up the search process.

Summary updated

BibTeX
@inproceedings{ye2018active,
  author = {Xin Ye and Zhe Lin and Haoxiang Li and Shibin Zheng and Yezhou Yang},
  title = {{Active Object Perceiver: Recognition-Guided Policy Learning for Object Searching on Mobile Robots}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2018},
  doi = {10.1109/iros.2018.8593720},
}
Figure