We update the installation guide for MinkowskiEngine compilation with CUDA 12.8. To compile MinkowskiEngine using CUDA 12.8, you should clone or update the source code and execute the following command before installation:
sed -i 's/\bauto __raw = __to_address(__r.get());/auto __raw = std::__to_address(__r.get());/' /usr/include/c++/11/bits/shared_ptr_base.hWe also support the new DINOv3 backbone as the dense encoder and simply test the performance on the "pour balls" task. The training and evaluation share the same settings with the experiments before (fix on Aug 29, 2025). During evaluation, we test the out-of-domain performance of DINOv2/v3 backbone in two levels:
- Level 1: novel cup and bowl 1.
- Level 2: novel cup, bowl 2 (smaller), and table.
 Train |
 Level 1 |
 Level 2 |
The parameter numbers and inference time are approximately the same for the two backbones. Average completion rates over ten scenes are as follows:
| Backbone | Level 1 | Level 2 |
|---|---|---|
| DINOv2 (base) | 0.9 | 0.45 |
| DINOv3 (base) | 0.8 | 0.8 |
The DINOv3 backbone generalizes better than the previous version on this task. Welcome to make comparisons on your own tasks and share the results with us.
We would like to thank Xiyan Huang for carrying out the experiments.
Due to unfixed point numbers of each point cloud in a batch, we apply WeightedSpatialInterpolation to each sample iteratively. However, in the original implementation in paper, MLP layers are also included in this module, which adopts BatchNorm as the normalization function. Multiple forwarding of the same BN during one training step hinders the model stability and leads to failure in some extreme cases. After the fix, the model shows improved performance.
We test the performance of RISE, RISE-2 (paper) and RISE-2 (fixed) using the task "pour balls" which is similar to the setting in RISE. The models are deployed using Nvidia RTX 3090, predicting 50 steps in a chunk and executing the first 20 steps with the control frequency of 10Hz. No trajectory smoothing strategy is employed to clearly compare the raw performance. The models are tested under 10 scenes. The completion rates (percentage of the balls successfully poured into the bowl) are as follows:
| Scene ID | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | Average | Stable Execution |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| RISE | 1 | 1- | 1- | 1 | 1 | 1 | 1 | 0- | 0- | 1 | 0.8 | 0.6 |
| RISE-2 (paper) | 1 | 1 | 1 | 1 | 1 | 1 | 1- | 1 | 1 | 1 | 1 | 0.9 |
| RISE-2 (fixed) | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
Although the completion rates before and after the fix are the same, we observe more unstable executions (inaccurate reaching, sharp transition between two action chunks, etc.) from "RISE-2 (paper)" (clearly unstable scenes are noted as "-" in the table). We select the most unstable scene "scene 7" and test the performance:
| Attempt ID | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | Average | Stable Execution |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| RISE-2 (paper) | 0- | 0- | 1 | 0- | 0- | 1- | 1 | 1- | 0- | 0- | 0.4 | 0.2 |
| RISE-2 (fixed) | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
The visualization of scene 7 for two implementations (download from video|gif if you cannot load the video here):
Comparison of RISE-2 performance before and after the fix.
From the table and video above, we can clearly observe a more stable execution using the fixed version.
NOTE: If you have already trained RISE-2 models using the old implementation and have no time for re-training, you could call policy.spatial_aligner.interp.train() after policy.eval() during evaluation to temporarily improve the performance, but it is still a bit unstable compared to the fixed version.
We would like to thank Xinyuan Guan for carrying out the experiments.