Fine manipulation tasks, such as threading cable ties or slotting a battery, are notoriously difficult for robots because they require precision, careful coordination of contact forces, and closed-loop visual feedback. Performing these tasks typically requires high-end robots, accurate sensors, or careful calibration, which can be expensive and difficult to set up. Can learning enable low-cost and imprecise hardware to perform these fine manipulation tasks? We present a low-cost system that performs end-to-end imitation learning directly from real demonstrations, collected with a custom teleoperation interface. Imitation learning, however, presents its own challenges, particularly in high-precision domains: errors in the policy can compound over time, and human demonstrations can be non-stationary. To address these challenges, we develop a simple yet novel algorithm, Action Chunking with Transformers (ACT), which learns a generative model over action sequences. ACT allows the robot to learn 6 difficult tasks in the real world, such as opening a translucent condiment cup and slotting a battery with 80-90% success, with only 10 minutes worth of demonstrations. Project website: https://tonyzhaozh.github.io/aloha/
核心贡献 · Key contributions
一个低成本的双臂遥操作系统（ALOHA），使用现成机械臂和 3D 打印部件构建，用于精细操作。 A low-cost bimanual teleoperation system (ALOHA) for fine manipulation, built with off-the-shelf arms and 3D-printed parts.
基于 Transformer 的动作分块（ACT），一种新颖的模仿学习算法，通过预测动作序列减少复合误差。 Action Chunking with Transformers (ACT), a novel imitation learning algorithm that predicts action sequences to reduce compounding errors.
时间集成以平滑重叠的动作块，在不增加训练的情况下改善运动平滑性。 Temporal ensembling to smooth overlapping action chunks, improving motion smoothness without additional training.
条件变分自编码器训练以处理多模态人类演示，对于从噪声数据中学习至关重要。 Conditional VAE training to handle multi-modal human demonstrations, crucial for learning from noisy data.
在 6 个真实世界精细操作任务上实现 80-90%的成功率，仅需 10 分钟演示。 Demonstrated 80-90% success on 6 real-world fine manipulation tasks with only 10 minutes of demonstrations.
开源硬件和软件设计，确保可复现性和可访问性。 Open-source hardware and software design for reproducibility and accessibility.
局限 · Limitations
限于机器人有效载荷（750 克）和工作空间（跨度 1.5 米）内的任务。 Limited to tasks within the robot's payload (750g) and workspace (1.5m span).
需要针对每个任务从头训练；无法跨任务泛化。 Requires task-specific training from scratch; no generalization across tasks.
在低对比度物体（如黑色桌面上的黑色扎带）上性能下降。 Performance degrades on low-contrast objects (e.g., black cable tie on black table).
大块大小的动作分块降低了对突发变化的反应能力。 Action chunking with large chunk sizes reduces reactivity to sudden changes.
遥操作需要用户技能；高频控制（50Hz）对精度至关重要。 Teleoperation requires user skill; high-frequency control (50Hz) is necessary for precision.
论文章节 · Sections（共 18）
摘要Abstract
一、引言I Introduction
二、相关工作II Related Work
三、ALOHA：低成本开源双臂遥操作硬件系统III ALOHA: A Low-cost Open-source Hardware System for Bimanual Teleoperation
四、基于 Transformer 的动作分块IV Action Chunking with Transformers
四-A 动作分块与时间集成IV-A Action Chunking and Temporal Ensemble
四-B 人类数据建模IV-B Modeling human data
四-C ACT 实现IV-C Implementing ACT
五、实验V Experiments
五-A 任务V-A Tasks
五-B 数据收集V-B Data Collection
五-C 实验结果V-C Experiment Results
六、消融实验VI Ablations
六-A 动作分块与时间集成VI-A Action Chunking and Temporal Ensembling