Two Reinforcement Learning Papers Accepted by IROS 2026

We are pleased to announce that two papers led by our master’s students, Changyou Miao and Li Teng, have been accepted for presentation at IEEE/RSJ IROS 2026.

Both works address the challenge of robotic harvesting in complex, occluded environments using reinforcement learning, with a shared focus on sim-to-real transfer and hierarchical policy design. The papers demonstrate how learning-based approaches can effectively handle the contact-rich, sequential nature of agricultural tasks such as obstacle separation and fruit detachment.

Changyou Miao (first author) and colleagues propose a reinforcement learning framework that unifies the full harvesting pipeline—obstacle separation, detachment, and placement—as a sequential decision-making problem. A hierarchical architecture combines high-level Cartesian actions with a low-level impedance controller for stable interaction under uncertain contact conditions. A feasibility-first observation alignment principle and domain randomization ensure robust zero-shot transfer from simulation to a structurally different real robot.

Li Teng (first author) and colleagues propose VGPA, a hierarchical reinforcement learning framework with a vision-guided decision mechanism and a Progressive Adaptive Exploration Strategy (PAES). The high-level vision module improves option selection and accelerates policy convergence, while PAES enhances exploration efficiency and training stability during continuous control. The framework is specifically designed for vision-based obstacle separation in clustered strawberry environments.

Congratulations to both teams on this achievement! 🎉

中文版

两篇强化学习论文被IROS 2026录用

我实验室两位硕士研究生 苗长友 和 李腾 作为第一作者的论文，近日被机器人领域顶级会议 IEEE/RSJ IROS 2026 正式录用。

两篇论文均聚焦于复杂遮挡环境下机器人采摘的挑战，采用深度强化学习方法，共同关注仿真到真实迁移与分层策略设计，展示了学习方法在处理农业操作任务中接触丰富、序列化决策问题上的有效性。

苗长友（第一作者）及合作者提出了一种强化学习框架，将避障、采摘和放置统一建模为序列决策问题。分层架构结合了高层笛卡尔空间动作与低层阻抗控制，确保在不确定接触条件下的稳定交互。可行性优先的观测对齐原则与域随机化策略，实现了从仿真到真实机器人的零样本迁移。

李腾（第一作者）及合作者提出了 VGPA 分层强化学习框架，包含视觉引导决策机制与渐进自适应探索策略（PAES）。高层视觉模块优化了选项选择并加速策略收敛，PAES则在连续控制学习中提升了探索效率与训练稳定性，专门面向簇生草莓环境下的视觉引导避障分离任务。

祝贺两位同学及合作团队！ 🎉