Abstract
Active Object Search (AOS) is a novel task requiring an intelligent agent to navigate a 3D environment to locate a target object with minimal steps. We present a modular reinforcement learning framework that integrates a 3D object detector, a specialized state controller, and a cost-sensitive reward function. Our approach achieves efficient, human-like search behavior by balancing detection accuracy with movement economy.
Introduction
Traditional object detection systems are typically passive, assuming the target is already centered in a static image. In real-world robotics, however, objects are often occluded, distant, or entirely out of view. An "active" agent must reason about its environment and plan movements—such as rotating or moving closer—to bring the target into focus.
The core challenge in AOS is optimizing the trade-off between search accuracy and efficiency. An agent must find the object correctly without unnecessary exploration.
Methodology
Our reinforcement learning framework consists of three primary components:
- 3D Object Detector: Uses Frustum PointNets on RGB-D data to predict 3D bounding boxes.
- State Controller: Monitors detection confidence (IoU) to emit a "stop" signal only when certainty is high, significantly reducing termination errors.
- Cross-modal Action Planner: A ResNet-18 and LSTM-based "brain" that processes visual scenes and target queries to decide the next action.
To encourage efficiency, we employ a Cost-Sensitive Active Search Reward (CSASR). The agent is rewarded for successful detection but penalized for every action taken, forcing it to find the most direct path.
The AOS Benchmark
We introduced the AOS benchmark using the AI2-THOR platform, featuring 5,845 samples across 30 indoor scenes. The dataset includes increased randomness in object placement and room layouts, categorized into three difficulty tiers:
- Easy: Target is initially in view.
- Moderate: Target is nearby but out of view.
- Hard: Target is distant, requiring multiple complex actions.
Results and Discussion
The CSASR method significantly outperforms passive and standard active baselines. While a naive static detector fails in Moderate and Hard scenarios, CSASR maintains a 50.53% recall in the most difficult settings.
Key insights include:
- Explicit Termination: A dedicated state controller reduced termination errors from 46% to 10%.
- Cost-Sensitivity: Penalizing movement mimics human-like efficiency, avoiding redundant actions.
- 3D Spatial Reasoning: 3D detection is vital for understanding depth and scale, enabling better navigation than 2D-only approaches.
Conclusion
AOS represents a significant step toward practical robotic vision. Future work will focus on sim-to-real transfer, multi-room environments, and integrating search with complex downstream tasks like object manipulation.
Summary updated