Region conditioned task decomposition with local graph repair for long horizon embodied agents

Long-horizon indoor embodied tasks require effective high-level organization, low-level coordination, and failure recovery. Existing methods often use generic subgoals or regenerate plans after execution failures, leaving spatial constraints and local repair boundaries insufficiently modeled. This paper presents a structured integration of room/region-conditioned task decomposition, memory-constrained local repair, and execution-derived auxiliary supervision. Each task-graph node encodes a target room or functional region, object set, subtask type, dependency information, and termination condition. When execution is disrupted, the framework localizes the affected subgraph and applies constrained graph editing while preserving still-valid nodes. Execution-consistency signals and proxy node-credit targets provide localized supervision for decomposition and repair learning. Experiments on ALFRED and TEACh show that, on ALFRED, the method improves GC, PLW-SR, and PLW-GC, while its mean SR remains slightly below ICAL without a statistically significant difference. On TEACh, it obtains numerically higher GC and PLW-GC, while none of the differences from ICAL, including SR, is statistically significant. These results indicate that the framework primarily benefits intermediate goal progression and path efficiency rather than overall final task success. Overall, the method extends existing hierarchical planning and memory-based recovery through an explicit structured task-graph interface.