从人类演示中进行的模仿学习在机器人领域展现了令人印象深刻的表现。然而，大多数结果集中在桌面操作上，缺乏完成一般有用任务所需的移动性和灵巧性。在这项工作中，我们开发了一个系统，用于模仿需要全身控制的双臂移动操作任务。我们首先提出了 Mobile ALOHA，一个用于数据收集的低成本全身遥操作系统。它在 ALOHA 系统的基础上增加了移动基座和全身遥操作界面。利用 Mobile ALOHA 收集的数据，我们随后进行监督行为克隆，并发现与现有的静态 ALOHA 数据集共同训练可以提升移动操作任务的性能。对于每个任务，使用 50 个演示，共同训练可以将成功率提高高达 90%，使 Mobile ALOHA 能够自主完成复杂的移动操作任务，例如翻炒和盛出一只虾、打开双门壁柜存放沉重的烹饪锅、呼叫并进入电梯，以及使用厨房水龙头轻轻冲洗用过的平底锅。项目网站：https://mobile-aloha.github.io
Imitation learning from human demonstrations has shown impressive performance in robotics. However, most results focus on table-top manipulation, lacking the mobility and dexterity necessary for generally useful tasks. In this work, we develop a system for imitating mobile manipulation tasks that are bimanual and require whole-body control. We first present Mobile ALOHA, a low-cost and whole-body teleoperation system for data collection. It augments the ALOHA system with a mobile base, and a whole-body teleoperation interface. Using data collected with Mobile ALOHA, we then perform supervised behavior cloning and find that co-training with existing static ALOHA datasets boosts performance on mobile manipulation tasks. With 50 demonstrations for each task, co-training can increase success rates by up to 90%, allowing Mobile ALOHA to autonomously complete complex mobile manipulation tasks such as sauteing and serving a piece of shrimp, opening a two-door wall cabinet to store heavy cooking pots, calling and entering an elevator, and lightly rinsing a used pan using a kitchen faucet. Project website: https://mobile-aloha.github.io
核心贡献 · Key contributions
提出了 Mobile ALOHA，一种低成本全身遥操作系统，用于双臂移动操作数据收集。 Presents Mobile ALOHA, a low-cost whole-body teleoperation system for bimanual mobile manipulation data collection.
展示了与静态 ALOHA 数据集联合训练，在 50 次示范下可将移动操作成功率提升高达 90%。 Shows that co-training with static ALOHA datasets boosts mobile manipulation success rates by up to 90% with 50 demonstrations.
在打开柜子、呼叫电梯和冲洗锅等复杂任务上，仅用 20-50 次示范即实现超过 80%的成功率。 Achieves over 80% success on complex tasks like opening cabinets, calling elevators, and rinsing pans with only 20-50 demonstrations.
证明了联合训练提高了数据效率和跨任务及模仿学习方法的泛化能力。 Demonstrates that co-training improves data efficiency and generalization across tasks and imitation learning methods.
提供了一个开源、低成本的硬件平台（3.2 万美元），用于全身遥操作和自主操作。 Provides an open-source, low-cost hardware platform ($32k) for whole-body teleoperation and autonomous operation.
通过在 7 个多样化的移动操作任务上进行大量真实世界实验，验证了该系统。 Validates the system through extensive real-world experiments on 7 diverse mobile manipulation tasks.
局限 · Limitations
仅限于单任务模仿学习；机器人无法自主改进或探索新知识。 Limited to single-task imitation learning; robot cannot improve autonomously or explore new knowledge.
示范由专家操作员收集；未在高度次优或异构数据集上测试。 Demonstrations collected by expert operators; not tested on highly suboptimal or heterogeneous datasets.
硬件占地面积（90 厘米×135 厘米）可能对狭窄路径过大；固定臂高限制了进入低柜的能力。 Hardware footprint (90cm x 135cm) may be too large for narrow paths; fixed arm height limits access to low cabinets.
炒虾任务成功率仅 40%，由于示范有限和视觉挑战。 Cook Shrimp task success rate only 40% due to limited demonstrations and visual challenges.
与静态数据联合训练可能对基于检索的方法（如 VINN）无益，由于动作预测不匹配。 Co-training with static data may not benefit retrieval-based methods like VINN due to action prediction mismatch.
论文章节 · Sections（共 13）
摘要Abstract
1 引言1 Introduction
2 相关工作2 Related Work
3 移动 ALOHA 硬件3 Mobile ALOHA Hardware
4 与静态 ALOHA 数据的协同训练4 Co-training with Static ALOHA Data
5 任务5 Tasks
6 实验6 Experiments
6.1 协同训练提升性能6.1 Co-training Improves Performance
6.2 与 ACT、扩散策略和 VINN 的兼容性6.2 Compatibility with ACT, Diffusion Policy, and VINN
7 消融研究7 Ablation Studies
8 用户研究8 User Studies
9 结论、局限与未来方向9 Conclusion, Limitations and Future Directions