aloha deployment question

[zhang-runhao]

Hello, dear author.

I have been using your code and noticed that the inference speed on the ALOHA dual-arm robot is quite slow. After careful investigation, I found that the bottleneck comes from the model’s autoregressive generation of all action tokens. It takes approximately 0.6–1.2 seconds to generate one action chunk, which causes noticeable pauses during execution.

May I ask how you addressed this issue during real-robot deployment? Would it be possible for you to share the deployment code with me via email at 3424224609@qq.com?

Thank you very much!

[Robertwyq]

@zhang-runhao
Hi, inference speed was indeed one of the limitations of our method at the time. This becomes especially noticeable in dual-arm settings, where the number of action tokens can be relatively large.
Our acceleration strategy back then mainly relied on using vLLM for faster inference, together with controlling the size of the action chunks.