Beyond LLMs: A Neuro-Symbolic Framework for Commonsense Reasoning in Robotics

Domestic robots often receive instructions that are underspecified regarding everyday actions. Large language models can interpret varied language and draw on everyday knowledge, but their reasoning is not explicitly inspectable. Symbolic systems provide transparent knowledge representation and inference, but depend on potentially incomplete rules. This thesis investigates whether combining these approaches can preserve or improve LLM performance on embodied commonsense tasks while making decisions more interpretable. Two complementary neuro-symbolic pipelines are evaluated on RoboCSKBench, a benchmark covering five embodied commonsense tasks. Pipeline 1 investigates the partially automated construction of Answer Set Programming rules for the Tidy Up and Tool Usage tasks, combining commonsense graphs, SOMA-inspired annotations, LLM-generated rule proposals, and Clingo verification. Pipeline 2 covers all five tasks: an LLM extracts controlled facts at runtime, while SWI-Prolog performs reasoning using knowledge-base evidence or hand-crafted rules. Pure-LLM and pure-logic ablations are evaluated using three Mistral model sizes, and Pipeline 2 is transferred without modification to five out-of-distribution datasets. Results show that symbolic reasoning is beneficial only under specific conditions. It is most effective when symbolic components can retrieve relevant affordance or capability knowledge and apply constraints aligned with the task’s answer format. In such cases, explicit grounding can reduce the performance gap between larger and smaller language models. Conversely, for open-ended generation or tasks whose expected decision structure is not supported by the symbolic selector, the architecture may become redundant or harmful. Overall, the findings support a conditional view of neuro-symbolic reasoning: it can improve decision traceability without necessarily sacrificing performance, provided that representation, knowledge, and inference are properly aligned.