LLM-Guided Cobot Control With Audited Task Instructions

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Large language models (LLMs) can generate unsafe or unpredictable responses, posing risks when used to control collaborative robots (cobots) in real-world scenarios, potentially leading to dangerous or harmful actions.

Innovation Solution

Constrain LLMs to generate human-readable discrete tasks that are audited for safety before being implemented by cobots, ensuring the tasks are safe and predictable by using a set of intermediate instructions that can be checked by humans or automated systems.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If LLMs are used to generate tasks for cobots, then the cobot can perform complex objectives with natural language queries, but the generated tasks may be unsafe or unpredictable

Engineering Contradiction:

Improveability to perform complex objectivesVSAvoidsafety and predictability of tasks

Core Design Contradiction:

Adaptability or versatilityVSReliability

Solution Approach 1:

The patent introduces an intermediary auditing system between the LLM and the cobot execution. This intermediary component reviews and validates the generated tasks before they are executed, ensuring safety and predictability while preserving the LLM's ability to generate complex objectives through natural language queries.

Inventive Principle:

Principle #24Intermediary (Mediator)

2Productivity

If LLMs generate computer-readable code directly, then the cobot can implement objectives efficiently, but unsafe actions cannot be prevented

Engineering Contradiction:

Improveimplementation efficiencyVSAvoidunsafe actions

Core Design Contradiction:

ProductivityVSObject-affected harmful factors

Solution Approach 1:

The patent implements preliminary action by requiring an auditing step before the cobot executes any generated code. The audit mechanism checks for safety and predictability in advance, preventing harmful factors from being executed while maintaining the efficiency of direct code generation from LLMs.

Inventive Principle:

Principle #10Preliminary action

3Reliability

If human-readable discrete tasks are added as intermediate instructions, then safety auditing becomes possible, but the system complexity increases

Engineering Contradiction:

Improvesafety auditing capabilityVSAvoidsystem structure

Core Design Contradiction:

ReliabilityVSDevice complexity

Solution Approach 1:

The patent applies segmentation by breaking down the LLM's generated output into human-readable discrete tasks as intermediate instructions. This segmentation enables systematic auditing of each task component while maintaining the overall flow from natural language query to cobot execution, managing complexity through structured decomposition.

Inventive Principle:

Principle #1Segmentation

Data Source

PatentUS12496726B2Control and training of robots using large language models

Publication Date: 2025.12.16 COLLABORATIVE ROBOTICS

AI summary

A method may include sending a query to one of a cobot or an LLM, where the query may be formed to trigger the cobot to perform a target objective, and where the LLM may be constrained to respond to requests to cause the cobot to perform a given objective with a set of human-readable discrete tasks to accomplish the given objective. The method may also include receiving a target set of human-readable discrete tasks from the LLM, where the target set of human-readable discrete tasks may correspond to target computer-readable code that is configured to cause the cobot to perform the target objective. The method may additionally include auditing the target set of human-readable discrete tasks, and, based on the target set of human-readable discrete tasks passing the audit, authorizing the cobot to implement the target computer-readable code to perform the target objective.