Core Concepts
Recover, a neuro-symbolic framework, leverages ontologies, logical rules, and large language models to efficiently detect failures during robotic task execution and generate recovery plans in real-time, enhancing the reliability and adaptability of autonomous systems.
Abstract
The paper introduces Recover, a neuro-symbolic framework for online failure identification and recovery in robotic task execution. Recover integrates ontologies, logical rules, and large language models (LLMs) to enhance the ability of LLMs to generate recovery plans and decrease the associated costs.
The key highlights are:
Recover utilizes an ontology, OntoThor, to describe the AI2Thor simulator environment, including objects, actions, spatial relations, and potential failures. This symbolic representation enables the robot to map multi-modal data to the same representation, allowing simultaneous reasoning across all data types.
Recover employs a set of logical rules defined in the ontology to accurately detect failures during task execution. The rules cover various failure types, including agent failures, environmental failures, planning failures, preference violations, and safety issues.
When a failure is identified, Recover uses an LLM as a re-planner to generate a new plan to recover from the failure and complete the task. By incorporating the symbolic information from the ontology, Recover can steer the LLM-planner towards generating fewer hallucinations and ensuring the system operates within the confines of available objects and actions.
Recover operates in an online manner, detecting failures during task execution and generating a new plan based on the current environment conditions, without the need to observe the effects of the failure throughout the entire plan.
The experiments demonstrate that Recover's rule-based sub-goal verifier achieves 100% accuracy in failure detection, and its ontology-enhanced LLM re-planning pipeline outperforms a purely LLM-based approach in both success rate and cost-effectiveness.
Stats
Recover demonstrated a significant reduction in monetary costs compared to a purely LLM-based approach.
Quotes
"By incorporating symbolic information, we can steer the LLM-planner towards generating fewer hallucinations, thereby ensuring that the system operates within the confines of available objects and actions."
"Recover detects failures during task execution and generates a new plan based on the environment conditions at the moment of failure, without needing to observe the effects of the failure throughout the entire plan."