Ideas for Improving the Autoware Diffusion Planner #6399
Unanswered
SakodaShintaro
asked this question in
Ideas
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
-
It's been a month since the initial pull request for
autoware_diffusion_plannerwas merged, and we've recently followed up with a version update (autowarefoundation/autoware_universe#11093). We're excited to continue improving the Diffusion Planner in an open community.This issue outlines our future plans and proposed improvements for the
autoware_diffusion_planner. While we may not be able to implement everything, we want to share these ideas to foster discussion.Multi-Batch Inference
Currently, inference is limited to a batch size of 1. We plan to add support for multi-batch inference. Multiple generated trajectories will be evaluated by a selector to choose the optimal one in the future. (See #6292)
2025/08/15 : Merged autowarefoundation/autoware_universe#11197
Effective Use of Ego Pose History
Currently, the trajectory is generated using only the information from the current frame. While we've recognized that incorporating Ego Pose History can lead to smoother speed control, we've encountered a challenge. The current implementation, when combined with the Trajectory Optimizer, prevents the vehicle from smoothly accelerating from a standstill. We believe this issue can be resolved by adjusting the Trajectory Optimizer, and we plan to introduce this feature once we find a suitable solution.
Drivable Area Integration
The Diffusion Planner currently relies primarily on VectorMap or Tracked Objects information. We aim to enhance its capabilities by integrating information about the drivable area. By using obstacle point clouds or occupancy grid map, we can improve the planner's ability to handle various situations. This may require implementing a way to accept rasterized image input.
Adding Semantic Information to VectorMap
The current model uses each Lanelet as a simple polyline. We intend to enrich the model's input with more semantic information from the VectorMap. This could include details such as:
Turn Indicator Topic Output
The model currently generates a turn indicator, but it is not yet published as a dedicated topic. We plan to add this functionality.
2025/08/25 : autowarefoundation/autoware_universe#11264
Beta Was this translation helpful? Give feedback.
All reactions