Multi-Object Robot Navigation Using Semantic Maps Without Pre-Sequencing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object navigation methods, particularly in visual object navigation for assistive agents, struggle to generalize to unseen environments and effectively utilize prior knowledge or semantic context for locating specific target objects, often requiring pre-sequencing and lacking efficiency in navigating multiple objects simultaneously.
Innovation Solution
A system and method for multi-object tracking and navigation that employs a deep reinforcement learning framework, utilizing a semantic map created from RGB-D observations and odometry pose readings to identify target objects and compute a long-term goal, allowing the robot to navigate efficiently without pre-sequencing, by using an actor-critic network to determine navigational actions and reward functions for reaching multiple target objects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If Reinforcement Learning techniques are employed for mapping pixels directly to actions, then the assistive agent can navigate towards target objects, but it finds it difficult to generalize to previously unseen scenarios since it lacks to build a representation of the unseen environment
Solution Approach 1:
The system performs preliminary actions by building a semantic map representation of the environment before navigation tasks are executed. This semantic map is constructed in advance and stored, allowing the RL agent to leverage this pre-built environmental understanding when navigating to target objects in unseen scenarios, thereby improving generalization without sacrificing navigation accuracy
2Adaptability or versatility
If the assistive agent implicitly learns the navigation problem through recursive trials, then it can adapt to the environment, but it requires extensive trials and time to locate target objects
Solution Approach 1:
The system performs preliminary actions by pre-building a semantic map of the environment that encodes spatial relationships and object locations. This eliminates the need for extensive recursive trials during actual navigation tasks, as the agent can directly query the pre-constructed semantic map for efficient path planning and target localization, significantly reducing the time required for adaptation
3Measurement precision
If the assistive agent uses a trained Siamese Network to locate target objects by finding similarity between current observation and target observation, then it can identify specific objects, but it lacks in utilizing prior knowledge or semantic context to locate target objects in unseen environments
Solution Approach 1:
The system introduces a semantic map as an intermediary between the Siamese Network and the environment. The semantic map encodes prior knowledge and semantic context about the environment structure and object relationships. When locating target objects, the system queries this semantic map to guide the Siamese Network's similarity matching process, enabling accurate object location in unseen environments by leveraging the pre-stored semantic context
4Measurement precision
If the system navigates to multiple target objects sequentially, then it can locate each object accurately, but it increases the total navigation time and reduces efficiency
Solution Approach 1:
The system dynamically adjusts the navigation plan by utilizing the pre-built semantic map to compute optimized paths that visit multiple target objects in an efficient sequence. Instead of rigid sequential navigation, the semantic map enables dynamic path re-planning that considers spatial relationships between multiple targets, allowing the agent to navigate to multiple objects with the same location accuracy but significantly reduced total navigation time
Data Source
Figure 1
Figure 2
Figure 3A
AI summary
This disclosure relates generally to method and system for multi-object tracking and navigation without pre-sequencing. Multi-object navigation is an embodied AI task where object navigation only searches for an instance of at least one target object where a robot localizes an instance to locate target objects associated with an environment. The method of the present disclosure employs a deep reinforcement learning (DRL) based framework for sequence agnostic multi-object navigation. The robot receives from an actor critic network a deterministic local policy to compute a low-level navigational action to navigate along a shortest path calculated from a current location of the robot to the long-term goal to reach the target object. Here, a deep reinforcement learning network is trained to assign the robot with a computed reward function when the navigational action is performed by the robot to reach an instance of the plurality of target objects.