HKUST PhD Chronicle, Week 16, Back to Deep Learning

My work on the Franka Research 3 is wrapping up. This week, I began a new task: using point clouds to identify a Teris block with full 6D pose estimation. The first paper I’m attempting to reproduce is PPR-Net++: Accurate 6-D Pose Estimation in Stacked Scenarios.

It’s been about 5 years since I last used PyTorch, back in graduate school when I applied graph neural network (primarily PyG) to predict textile shape deformation. Returning to the material feels a bit unfamiliar. I even found myself hesitating over how to construct a simple torch.tensor.

data = [[1, 2], [3, 4]]
x_data = torch.tensor(data)

Still, I am now on board with this new challenge. I’m hoping to uncover something interesting along the way.

See those T Tetris blocks? Yup, that’s my current goal: predicting each one’s 6D pose from point cloud data.