search

YanjieZe / 3D-Diffusion-Policy

[RSS 2024] 3D Diffusion Policy: Generalizable Visuomotor Policy Learning via Simple 3D Representations
https://3d-diffusion-policy.github.io
MIT License
402 stars 37 forks source link

Question Regarding Ablation Study of the Encoder #39

Closed liops-seungyeob closed 3 months ago

liops-seungyeob commented 3 months ago

Since the input size of the Diffusion Policy is fixed, the output of the Point Cloud Encoder must also consistently return a fixed value.

It appears that the DP3 and PointNet encoders use Max Pooling to achieve this.

For other encoders, what methods were used to match the input size of the Diffusion Policy?

YanjieZe commented 3 months ago

Hi, thank you for your interest. For other encoders, they all output a latent vector for downstream tasks, so it is similar.