BAKU: An Efficient Transformer for Multi-Task Policy Learning

训练能够解决各种任务的通用代理人是具有挑战性的，通常需要大量专家演示数据集。在机器人领域，这尤其成为问题，因为每个数据点都需要在现实世界中执行物理动作。因此，需要一种能够有效利用可用训练数据的架构。在本研究中，我们提出了BAKU，一种简单的Transformer架构，能够有效地学习多任务机器人策略。BAKU建立在离线模仿学习的最新进展基础上，精心组合了观察框架、动作分块、多感官观察和动作头，从而大大改进了之前的工作。我们在LIBERO、Meta-World套件和Deepmind Control套件的129个模拟任务上进行的实验表明，相对于RT-1和MT-ACT，我们的方法整体上有18%的绝对改进，LIBERO基准测试的改进率为36%。在30个真实世界的操作任务中，BAKU在每个任务平均仅17个演示的情况下实现了91%的成功率。机器人的视频最好在https://baku-robot.github.io/观看。

BAKU: Efficient Multi-Task Robot Learning from Offline Demonstrations

BAKU is a transformer architecture that enables efficient learning of multi-task robot policies by combining observation trunks, action chunking, multi-sensory observations, and action heads to substantially improve upon prior work.

BAKU achieves an overall 18% absolute improvement over RT-1 and MT-ACT on 129 simulated tasks across LIBERO, Meta-World suite, and the Deepmind Control suite, with a 36% improvement on the harder LIBERO benchmark. On 30 real-world manipulation tasks, given an average of just 17 demonstrations per task, BAKU achieves a 91% success rate. The robot videos are available on the project website.

Recent related work includes 'RT-1: Efficient Multi-Task Reinforcement Learning with Successor Features and Temporal Abstraction' and 'MT-ACT: Efficient Multi-Task Reinforcement Learning with Model Truncation'.