Semantic Robot Navigation With Trajectory History for Multi-Object Scenes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional semantic navigation methods for service robots in indoor environments struggle to effectively navigate to out-of-view objects, especially when multiple objects are present in the same scene, due to limitations in decision-making models and the requirement for preset layouts and prior object-object relations.
Innovation Solution
A method and system for semantic navigation using a spatial graph and trajectory history, which involves receiving a target object, identifying the current location of the mobile robot, computing embeddings for visible regions using a pretrained Graph Neural Network (GNN), calculating similarity scores, and selecting the next action based on the optimal visible region.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional GNNs based methods are used for semantic navigation, then the robot can navigate in indoor environments, but the decision model fails when two or more objects are in the same scene
Solution Approach 1:
The patent segments the navigation problem into multiple components: spatial graph construction for environmental structure, trajectory history for temporal context, and region-based embedding for spatial relationships. This segmentation allows the system to handle multiple objects by considering their relative positions and trajectories separately, resolving the decision model failure in multi-object scenes.
Solution Approach 2:
The patent introduces temporal dimension through trajectory history and spatial dimension through region embeddings, transforming the navigation decision from a single-state problem to a multi-dimensional problem. This allows the system to distinguish between multiple objects by considering their positions across different time steps and spatial regions, improving reliability in multi-object scenarios.
2Adaptability or versatility
If BRM model is used for semantic navigation, then generalization ability is improved for unseen environments, but preset layout and prior object-object relations are required
Solution Approach 1:
The patent performs preliminary actions by constructing spatial graphs and computing region embeddings during the exploration phase. This preliminary spatial understanding is stored and reused during navigation to unseen environments, achieving generalization without requiring preset layouts. The spatial relationships are learned dynamically rather than being pre-defined.
Solution Approach 2:
The system serves itself by automatically constructing spatial representations and learning object relationships through trajectory data. Instead of requiring external preset layouts and prior knowledge, the system builds its own spatial understanding from raw sensor data and movement history, eliminating the need for complex preset configurations.
3Productivity
If service robots navigate to out-of-view objects in indoor environments with multiple object instances, then navigation task is completed, but conventional methods lack concrete decision model
Solution Approach 1:
The patent introduces region embeddings as an intermediary representation between raw sensor data and navigation decisions. These embeddings capture spatial relationships and object contexts, serving as a mediator that simplifies the decision-making process. Instead of directly handling complex multi-object scenarios, the system uses region embeddings to represent the environment, making navigation decisions more tractable and concrete.
Solution Approach 2:
The system uses trajectory history as feedback to improve navigation decisions. By continuously updating the spatial graph with observed trajectories and using this historical information to inform future decisions, the system develops a concrete decision model that adapts to the environment. This feedback mechanism enables the robot to learn from past experiences and make more reliable decisions in multi-object scenarios.
Data Source
AI summary
A model for semantic navigation for service robots to find out-of-view objects in an indoor environment is provided. Initially, the system receives a target object to be reached by the mobile robot in the indoor environment. Further, a current location of the mobile robot is identified by a localization technique. An embedding corresponding to each of a plurality of visible regions is computed using a pretrained Graph Neural Network GNN. The GNN is pretrained using a trajectory data and a spatial relationship graph associated with the indoor environment. Further, a similarity score is computed for each of the plurality of visible regions based on the corresponding embedding using a scoring technique. An optimal visible region is identified by comparing the similarity score. Finally, a next action to be taken by the mobile robot selected from a plurality of actions based on the optimal visible region.