Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (ALOHA)

Main contributions of the paper:

  • Introducing the concept of action chunking to reduce the compounding error problem

“Addressing compounding errors. A major shortcoming of BC is compounding errors, where errors from previous timesteps accumulate and cause the robot to drift off of its training distribution, leading to hard-to-recover states

  • This is due to the autoregressive nature of the models

  • So the trajectory goes more and more out of distribution

That’s why Action chunking can help, so it reduces the effective horizon of the task by “k”-fold.

ACT is trained as a cVAE