[Question] Action decoder

[zhyan330]

In the TrackVLA paper, a Diffusion Transformer is used as the action output head. However, in this implementation, it has been replaced with an MLP. Besides ease of implementation, are there any other advantages? Has there been a comparison of performance between these two types of output heads?

[kyusonglee]

We explored using a diffusion transformer for the action head, but couldn't reproduce the quality reported in the original TrackVLA paper. The trajectories didn't look good enough. We switched to an MLP-based approach instead, which proved to work well. The main advantage is its simplicity.

[zhyan330]

Thanks for your reply. I’ve recently successfully run the training and deployment pipeline based on the code you open-sourced. I really appreciate your work! However, I encountered two issues during practical use: 1. *The model currently lacks sufficient target recognition capability.* Specifically, the presence of interfering people affects the tracking of the target person. 2. *The model seems to rely heavily on the current frame's image for tracking.* Once the target disappears from the current frame, the model is unable to continue tracking. Do you have any good suggestions regarding these two issues? Also, we noticed that you released a new version of model weights a couple of days ago. Could you please share what specific modifications or improvements contributed to the performance boost in this update? Looking forward to your reply Kyusong Lee ***@***.***> 于2026年4月15日周三 22:24写道：

…

Closed #20 <#20> as completed. — Reply to this email directly, view it on GitHub <#20 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BX5JREEMCRRFWMN3W3QRW334V6LTJAVCNFSM6AAAAACXFXNRG2VHI2DSMVQWIX3LMV45UABCJFZXG5LFIV3GK3TUJZXXI2LGNFRWC5DJN5XDWMRUGUZTEMZQGQ4TKNQ> . You are receiving this because you authored the thread.Message ID: ***@***.***>