Robot Object Grasping with Confidence-Based Action Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Service robots face challenges in executing tasks when encountering untrained instructions or objects with similar attributes, leading to low success rates and efficiency in task execution.
Innovation Solution
An object grasping method that determines the confidence of candidate objects being the target object based on state description information and preset states, and performs preset actions such as grasping and questioning to determine the optimal action for object grasping.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the robot executes pre-trained tasks only, then the task execution reliability is improved for known tasks, but the adaptability deteriorates when untrained tasks or instructions occur
Solution Approach 1:
The patent implements a dynamic task execution framework that transitions between pre-trained task execution and active learning modes. The system dynamically adjusts its operation based on whether the current task is recognized as pre-trained or novel, enabling both high reliability for known tasks and adaptability for untrained tasks through real-time mode switching
Solution Approach 2:
The robot performs self-learning by actively asking questions when encountering untrained tasks. The system uses its own curiosity-driven questioning mechanism to gather information about novel tasks, enabling it to independently expand its capability without external intervention, thus improving adaptability while maintaining reliability
2Measurement precision
If the robot identifies objects one by one in scenarios with similar attributes, then the measurement precision is improved for individual object identification, but the productivity deteriorates due to increased identification time
Solution Approach 1:
The patent segments the object identification process into two stages: first, group objects with similar attributes using clustering algorithms to reduce the search space; second, apply precise identification methods only within each cluster. This segmentation maintains high identification precision while significantly improving efficiency by avoiding sequential comparison of all objects
Solution Approach 2:
The system applies partial identification actions by first performing attribute-based grouping rather than complete individual identification for all objects. This partial action approach identifies objects efficiently in groups, reserving detailed individual identification only when necessary, thus balancing precision and productivity
3Measurement precision
If the robot asks more questions to clarify user instructions, then the task execution accuracy is improved, but the loss of time increases due to extended conversation duration
Solution Approach 1:
The patent performs preliminary analysis of user instructions using pre-trained language models to predict the most likely intended tasks before engaging in conversation. This preliminary action allows the system to proactively ask targeted clarification questions rather than engaging in extended back-and-forth dialogue, reducing conversation time while maintaining accuracy
Solution Approach 2:
The system implements feedback mechanisms where each question asked is based on previous interaction outcomes and confidence level assessments. The robot continuously updates its understanding based on user responses and adjusts subsequent questions accordingly, minimizing unnecessary conversations and reducing time loss while maintaining high instruction understanding accuracy
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
The present disclosure provides an object grasping method, an apparatus, a computer device and a storage medium, the method includes: in response to receiving a target object grasping instruction, determining a second confidence of each of candidate objects being a target object based on state description information of the target object and a first confidence of each of the candidate objects in a current scene having each preset state; determining a target candidate object and a reward information after performing each preset action for the target candidate object based on the first confidence and the second confidence; and determining a target preset action to be performed based on the reward information, and performing the target preset action. In the embodiments of the present disclosure, the robot does not need to be pre-trained, and the target candidate object that meets the requirements of the user can be determined more accurately; and by determining the optimal target preset action, the number of conversations with the user is reduced, the target object can be grasped as soon as possible, and the success rate and efficiency of task execution are improved.