BAKU: An Efficient Transformer for Multi-Task Policy Learning

训练能够解决各种任务的通用智能体具有挑战性，通常需要大量专家演示的数据集。这在机器人技术中尤为棘手，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现有训练数据的架构。在这项工作中，我们提出了BAKU，这是一种简单的变压器架构，可以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之上，并精心组合了观察树干、动作分块、多感官观察和动作头，从而大大改善了以前的工作。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，总体上实现了18%的绝对改进，在更难的LIBERO基准测试中实现了36%的改进。在30个真实世界的操作任务中，每个任务平均只给出17个演示，BAKU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 via Reward Prediction

BAKU is a transformer architecture that enables efficient learning of multi-task robot policies by meticulously combining observation trunks, action chunking, multi-sensory observations, and action heads.

BAKU exhibits 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 at https://baku-robot.github.io/.

Recent related research includes 'Off-Policy Reinforcement Learning for Robotic Manipulation with Latent Space Planning' and 'Learning Dexterous In-Hand Manipulation'.