Abstract
Autonomous systems possess the features of inferring their own ego-motion, autonomously understanding their surroundings, and planning trajectories. With the applications of deep learning and reinforcement learning, the perception and decision-making abilities of autonomous systems are being efficiently addressed, and many new learning-based algorithms have surfaced with respect to autonomous perception and decision-making. In this review, we focus on the applications of learning-based approaches in perception and decision-making in autonomous systems, which is different from previous reviews that discussed traditional methods. First, we delineate the existing classical simultaneous localization and mapping (SLAM) solutions and review the environmental perception and understanding methods based on deep learning, including deep learning-based monocular depth estimation, ego-motion prediction, image enhancement, object detection, semantic segmentation, and their combinations with traditional SLAM frameworks. Second, we briefly summarize the existing motion planning techniques, such as path planning and trajectory planning methods, and discuss the navigation methods based on reinforcement learning. Finally, we examine the several challenges and promising directions discussed and concluded in related research for future works in the era of computer science, automatic control, and robotics.
Introduction
Traditional Visual Simultaneous Localization and Mapping (vSLAM) has long served as the backbone of robotic autonomy. However, classical methods typically rely on hand-crafted features and the "static world" assumption, which often fail in dynamic, textureless, or low-light environments. The transition into the "Era of Learning" addresses these limitations by enabling robots to move beyond simple geometric point clouds toward a more human-like understanding of their surroundings. This shift involves using Deep Learning (DL) for sophisticated environmental perception and Reinforcement Learning (RL) for intelligent, end-to-end decision-making.
System Architecture
Modern autonomous systems are structured to bridge the gap between raw sensor data and intelligent action. A typical architecture involves a closed-loop interaction between the agent and its environment, mediated by perception and planning modules.
As shown in the diagram, the Perceiver processes environmental data and agent state, feeding information to both the Motion Planning Module (which considers the mission goals) and the Controller. The controller then issues commands to the Agent, which acts upon the Environment, completing the loop.
Learning-Based Perception
Perception serves as the "eyes" of the system. Learning-based approaches have revolutionized how robots represent their environment, moving from sparse geometric points to dense, semantically rich maps.
- Monocular Depth Estimation: CNNs now allow for 3D structure estimation from a single 2D image, reducing reliance on expensive LiDAR. Both supervised and unsupervised methods (leveraging geometric constraints like smoothness loss) have shown significant progress.
- Ego-Motion Prediction: Networks like PoseNet and DeepVO directly regress camera movement from image sequences, bypassing traditional feature-matching bottlenecks.
- Image Enhancement: Generative Adversarial Networks (GANs) can translate challenging inputs—such as night-time or foggy scenes—into clear, "day-time" versions, extending the operational range of standard SLAM algorithms.
- Semantic Integration: By incorporating object detection (e.g., YOLO) and semantic segmentation (e.g., Mask R-CNN), systems can identify and ignore dynamic objects like moving cars, which would otherwise introduce noise into the localization process.
Decision-Making and Navigation
Decision-making acts as the "brain," translating perceived information into trajectories. The field is moving toward more flexible, learning-based navigation strategies.
- Deep Reinforcement Learning (DRL): Algorithms like DQN and DDPG enable mapless navigation, where agents learn optimal paths through trial and error in simulation. This allows for better generalization to unseen environments compared to rigid traditional planners.
- Visual Language Navigation (VLN): This emerging field enables agents to follow natural language instructions (e.g., "Go to the kitchen") by aligning linguistic cues with visual perception, moving closer to intuitive human-robot interaction.
Key Insights and Future Directions
The integration of learning has effectively challenged the "static world" assumption. By using semantic masks to filter dynamic noise, learning-based SLAM achieves superior accuracy in crowded urban settings. Furthermore, the ability of deep networks to learn environmental "priors" allows for accurate depth estimation from single lenses, rivaling stereo camera performance.
Despite these successes, several challenges remain:
- Sim-to-Real Gap: RL agents often struggle to transfer skills from high-fidelity simulators (like AI2-THOR) to the noisy real world.
- Computational Constraints: High-accuracy models require significant resources, posing a hurdle for real-time deployment on low-power mobile platforms.
- Interpretability and Safety: The "black box" nature of deep learning remains a concern. Future research is trending toward geometry-assisted perception, using known physical laws to constrain and explain neural network outputs.
In conclusion, the synergy between traditional geometric methods and modern learning-based approaches is paving the way for more robust, adaptable, and intelligent autonomous systems.
Summary updated