Long-horizon embodied manipulation: a survey of methods, trends, and challenges
Abstract Long-horizon embodied manipulation requires robots to translate abstract goals into extended sequences of reliable physical interactions under partial observability, contact dynamics, sparse feedback, and embodiment constraints. Its difficulty arises not merely from horizon length, but from maintaining semantic dependencies, physical feasibility, task progress, and recoverability across interdependent object-state transformations. This survey systematically reviews the methods, trends, and challenges of this field. We formalize long-horizon embodied manipulation across primitive-action, skill, interaction-event, and semantic-goal horizons, and organize existing methods around three coupled bottlenecks: task structuring, grounded interaction, and progress and adaptation. Accordingly, we examine language- and vision-language-based decomposition, hierarchical reinforcement learning, skill composition, action chunking, sequence and diffusion planning, world models, affordance grounding, progress and reward modeling, memory, replanning, and self-correction. Our analysis reveals a transition from open-loop, model-specific solutions toward hierarchical, affordance-aware, progress-sensitive, and closed-loop systems that coordinate reasoning and control. We further review embodied and simulation benchmarks, arguing that binary success obscures partial completion, error propagation, recovery cost, and horizon-dependent degradation. Finally, we identify open challenges in subgoal reachability, planner-controller alignment, irreversible failure, transferable progress supervision, real-time adaptation, cross-embodiment generalization, sim-to-real reliability, and scalable evaluation, providing a unified foundation for developing dependable robots capable of sustained manipulation in complex environments.
Authors
- Rufan Bai (ORCID: https://orcid.org/0009-0009-9059-8141)
- Yuhuan Lu (ORCID: https://orcid.org/0000-0001-5332-3389)
- Binbin Song (ORCID: https://orcid.org/0000-0002-2448-1932)
- Yapeng Wang
- Qifan Li
- Yi Zheng
- Kui Jiang
Institutions
- Harbin Institute of Technology (CN)
- Suzhou Research Institute (CN)
- Macao Polytechnic University (MO)
- Southeast University (CN)
Publication Details
- Journal
- Vicinagearth.
- Published
- 2026-11-16
- DOI
- https://doi.org/10.1007/s44336-026-00048-x
- Primary Topic
- Robot Manipulation and Learning
- Type
- article
- Field-Weighted Citation Impact
- 0.00