ELEVATE: Agentic Controller Acquisition from Task Failure for Quadrupedal Manipulation

When a quadruped manipulator keeps failing, a coding agent writes an RL problem to learn the missing whole-body skill, adds it to a library and retries; transfers to a real robot (18/20 rack retrieval).四足机械臂反复失败时，由编码智能体判断缺少哪种全身控制能力，自动编写强化学习问题（奖励、课程、终止条件）训练新控制器并加入技能库再重试；习得扶墙抬升等技能，仿真五项超出初始能力的任务成功率 98.4–100%，无需真机训练即迁移到实体机器人（货架取物 18/20）。

@misc{liu2026elevate,
  title = {{ELEVATE: Agentic Controller Acquisition from Task Failure for Quadrupedal Manipulation}},
  author = {Shiqi Liu and Radu Corcodel and Ding Zhao},
  year = {2026},
  eprint = {2610.38388},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  url = {https://arxiv.org/abs/2610.38388},
}