Abstract: Long-horizon robot manipulation is a critical bottleneck in transitioning embodied intelligence from short-horizon skill execution to real-world deployment. It is not merely an extension of action sequences, but a sequential embodied decision-making problem jointly constrained by stage structure, history dependence, and error propagation. Unlike short-horizon manipulation, which primarily relies on current observations and local feedback, long-horizon manipulation requires a robot to maintain a global objective across multiple interdependent stages. In household service, manufacturing assembly, logistics, mobile manipulation, and contact-rich scenarios, early actions may alter the execution preconditions and reachable sets of later stages, intermediate states may become partially observable, and local deviations may accumulate into cascading failures. Existing surveys encompass embodied intelligence, robotic foundation models, vision-language-action models, world models, and task planning. However, they are commonly organized by model type and therefore provide limited explanations of why short-horizon policies fail in long-horizon tasks, how failure modes arise from task structure, and how different mechanisms compensate for these structural difficulties. To address this gap, this study adopts a task-structure-centered perspective and develops a unified “problem–challenge–mechanism–validation” framework to organize representative studies. Specifically, cross-stage conditional dependence is taken as the basic criterion for long-horizon manipulation, and task properties and difficulty are characterized along three coupled dimensions: stage structure, history dependence, and error propagation. Stage structure describes how a global task is organized into multiple subgoals connected by precedence relations, completion verification, and possible fallback paths. History dependence means that current observations may be insufficient to determine task progress or object state, requiring the integration of past observations, actions, stage feedback, and semantic context. Error propagation describes how local deviations alter the execution preconditions and reachable sets of subsequent stages to increase the recovery scope and cost. Based on this characterization, a unified framework is established to map structural difficulties to compensatory mechanisms. Three core challenges are identified: combinatorial explosion and path expansion, state drift under partial observability, and error propagation with increasing recovery cost. Existing methods are organized into four functional mechanisms: task decomposition and planning, world models and forward prediction, state and semantic support, and closed-loop control with online correction. These mechanisms support stage organization, future-state rollout and reachability assessment, cross-stage context maintenance, and hierarchical execution recovery, respectively. Their coordination is essential because improvements in local control accuracy alone cannot guarantee consistency across stages, preserve task context, or prevent error accumulation over extended execution. The study further examines the formation, evaluation, and deployment of long-horizon capabilities. Training should extend from one-step imitation to stage-level supervision, trajectory consistency, skill-progress modeling, and recovery learning. Evaluation should move beyond final success rate toward process-oriented indicators, including stage completion, progress maintenance, planning-execution consistency, context maintenance, and recovery performance. Deployment should account for cloud-edge-device collaboration, latency, privacy, safety loops, and physical failure costs. Overall, reliable long-horizon manipulation depends on the coordinated maintenance of task progress, historical context, future reachability, and recoverable execution under partial observability and accumulated errors. This perspective provides a structured basis for comparing methods, designing benchmarks, and developing robust, interpretable, and recoverable embodied systems for realistic and dynamically changing environments.