Event

CausGround: Causal Interaction Knowledge Improves Embodied Planning Success Rate

1 reports · 1 sources

Yijun Shen et al. propose CausGround, a non-parametric causal interaction grounding framework that distills interaction experience into action-centric reusable causal knowledge, identifies action preconditions and effects and organizes them into a decomposed world-state representation, aligns them with current observations during planning, and supports backward reasoning over preconditions and forward reasoning over effects. On modified VirtualHome and RoboCasa benchmarks, CausGround improves visual high-level planning performance across multiple multimodal large-model backbones: VirtualHome average task success rate rises from 24.9% to 53.3%, RoboCasa closed-loop planning success rate rises from 43.0% to 57.0%; on SWITCH, average interaction reasoning accuracy rises from 25.8% to 31.9%. Real-robot experiments show that CausGround can backward-derive executable high-level action sequences from goals interpreted by large models for sequential manipulation tasks, without updating model parameters.