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

Led by Mr. Tony Zhao.

• https://tonyzhaozh.github.io/aloha/
• https://mobile-aloha.github.io/
• https://aloha-2.github.io/
• https://aloha-unleashed.github.io/

Papers

• https://arxiv.org/abs/2304.13705

Related

• UMI

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

Links

• https://tonyzhaozh.github.io/aloha/
• paper https://arxiv.org/pdf/2304.13705

Resources

• https://radekosmulski.com/how-to-train-your-robot-with-a-transformer/

So what is action chunking? Action chunking is a technique in robotics and imitation learning where, instead of predicting and executing one action at a time, a robot forecasts and carries out a sequence of actions—referred to as a “chunk”—based on its current observation.

Benefits of action chunking (reference blog):

• Allows your model to control your robot at a much higher frequency given a large model
  • You of course don’t get the reactivity, but that’s a bonus
• Better temporal consistency without having to do Proprioceptive History

“Generally, the more information is contained in the observation space, the more likely causal confusion will happen for imitation learning.”

Related

• Real-Time Execution of Action Chunking Flow Policies

Citation

(Zhao et al., 2023)

Zhao, T. Z., Kumar, V., Levine, S., & Finn, C. (2023). Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. https://arxiv.org/abs/2304.13705