Robotic Motion Planning From Natural Language in Continuous Spaces
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing approaches to task and motion planning in robotics rely on symbolic representations, which limit their ability to handle complex tasks, learn new concepts, and navigate ambiguous environments.
Innovation Solution
A method that combines task and motion planning using a neural network and a planner, operating in continuous action and state spaces without symbolic representations, allowing for the execution of natural language commands in complex environments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If symbolic representations are used in high-level planning languages, then task planning and motion planning can be combined, but the system cannot learn new concepts and cannot handle ambiguity well
Solution Approach 1:
The patent replaces symbolic representation mechanisms with neural network-based continuous representation. Instead of using discrete symbolic states and actions, the system employs continuous state vectors and action vectors that are processed by neural networks, enabling the system to learn new concepts from data without requiring manual symbolic definitions.
Solution Approach 2:
The patent changes the parameter space from discrete symbolic values to continuous numerical values. State and action are represented as continuous vectors rather than discrete symbols, allowing the system to generalize to new situations and learn new concepts by interpolating in the continuous space rather than requiring explicit symbolic definitions.
2Adaptability or versatility
If symbolic representations are used for task planning, then planning in large continuous spaces is possible, but manual creation of symbolic representations is required
Solution Approach 1:
The patent implements self-service by enabling the system to automatically learn its own representations from data. The neural networks are trained on interaction data to automatically discover useful state and action representations, eliminating the need for manual creation of symbolic representations while maintaining the ability to handle complex environments.
Solution Approach 2:
The patent performs preliminary learning actions during a training phase where the neural networks are trained on collected interaction data. This preliminary action of learning representations from data prepares the system to automatically handle complex environments without requiring manual symbolic representation creation during operation.
3Ease of operation
If continuous action and state spaces are used, then natural language commands can be followed, but many precise steps are required to execute single output tokens
Solution Approach 1:
The patent segments the complex continuous planning problem into manageable components by using a neural network to predict the next action state based on current state and natural language input. This segmentation allows the system to process natural language commands step-by-step through sequential predictions rather than requiring all precise steps to be predetermined.
Solution Approach 2:
The patent introduces neural networks as intermediaries between natural language input and continuous action output. The neural network acts as a mediator that translates natural language tokens into continuous action vectors, bridging the gap between high-level language commands and low-level continuous control without requiring manual specification of every precise step.
Data Source
AI summary
The present approach similarly combines task and motion planning, but does so without symbolic representations and begins with simpler tasks than other models in such domains can handle. Unlike prior approaches, the present approach does so in continuous action and state spaces which require many precise steps in the configuration space to execute what otherwise is a single output token such as “pick up” for discrete problems.