Johnson Zhong

Can confirm this also happens when adding new runs.

Hi, what is the input `obs` that you're feeding into `ctrl.command`? It should be a `torch.tensor` or convertible type (such as list, or numpy ndarray) of size `nx` or `K...

I see, this seems to be a relatively big problem since even adding noise didn't resolve tie-breaking (tried rand up to magnitudes up to 1).

I see, I tried this again and the critical line is decreasing the temperature. There is a new issue however with some hallucinated values such as: ``` 42.00003 , 84.00015...

Thanks, will keep in mind!

The spurious values seems to be from `IndexedMatmul2Efficient` in aggregating the output (produce z from W) instead of from the `log_softmax`. For example, the W of querying for k=3 nearest...

Fixing this by normalizing each output distribution removes the spurious values, but still ends up selecting the center value multiple times because of aggregation via expected value. It seems like...

@bottler Yes there are a couple, including the NICP (with implementation https://laempy.github.io/pyoints/tutorials/icp.html#NICP-algorithm; J. Serafin and G. Grisetti (2014): ‘NICP: Dense normal based pointcloud registration’, International Conference on Intelligent Robots and...

I am currently investigating this problem and seeing how well existing solutions handle some related conditions (3D NDT and SDF to SDF registration methods), as well as developing my own....

https://youtu.be/h3zAA3p2A-I preliminary results on YCB mustard bottle registration with ICP using 5 known surface points (shown as black crosses, with a blue line coming out of them indicating their normal)....