Robot Trajectory Generation Using Success-Failure State Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for designing controllers for compliant robots face challenges due to complex dynamics, manual task division, variable performance based on failure types, and simulation limitations, making it difficult to achieve high success rates in tasks like peg insertion.
Innovation Solution
A trajectory generation device and method that utilizes successful and failed trajectory information to generate target trajectory information using a Gaussian mixture model for clustering and regression, adjusting similarity degrees to separate successful and failed states, and generating target trajectories.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If reinforcement learning is used for controller design, then automation is improved, but difficulty of detecting and measuring increases due to complex dynamics
Solution Approach 1:
The patent introduces an intermediary system consisting of sensors, state estimation algorithms, and data processing modules that mediate between the complex robotic system and the reinforcement learning controller. This intermediary layer simplifies the dynamics by providing processed state information, making the control problem more tractable while maintaining automation.
2Manufacturing precision
If manual task division is used, then manufacturing precision is improved, but device complexity increases
Solution Approach 1:
The patent segments the control task into hierarchical levels: high-level task planning and low-level execution control. This segmentation allows precise control at each level while avoiding the complexity of designing a single monolithic controller, as the system divides responsibilities between different control modules.
Solution Approach 2:
The patent implements dynamic task decomposition where the division of tasks is not fixed but adapts based on real-time system state and performance. This dynamic approach allows the controller to automatically adjust the granularity and allocation of sub-tasks, maintaining precision without requiring complex pre-programmed task structures.
3Measurement precision
If reward functions are designed to follow target states precisely, then measurement precision is improved, but reliability decreases due to excessive force and overshooting
Solution Approach 1:
The patent dynamically changes the parameters of the reward function based on the task phase and system state. Instead of using a fixed precision-based reward, the system adjusts weighting factors and target tolerances according to compliance requirements, contact status, and task progress. This allows the controller to achieve sufficient position accuracy while avoiding excessive forces and overshooting that would reduce reliability.
Data Source
AI summary
A trajectory generation device includes an acquisition unit, a clustering unit and a generation unit. The acquisition unit acquires successful trajectory information and failed trajectory information that are trajectory information representing sequences of states of a controlled object being taught by a teacher. The successful trajectory information is trajectory information when a task performed by the controlled object is successful, and the failed trajectory information is trajectory information when the task is failed. From the states of the controlled object belonging to the successful trajectory information and the states belonging to the failed trajectory information, the clustering unit generates clusters of successful classes of the states. On the basis of the clusters of successful classes, the generation unit generates target trajectory information representing a sequence of states of the controlled object, the sequence of states being usable as control targets when the controlled object is caused to execute the task.