2025.05.15 - #35 - SceneScript, GLIM-CPU, Google's 3D generation, PyRoki, VINS-Multi, ForceMimic

[changh95]

Interesting papers

Building Rome with Convex Optimization

• XM: new structure from motion pipeline powered by GPU-accelerated convex semidefinite optimization:
• Large scale global bundle adjustment via learned depth & tight convex semi-definite relaxation
• https://github.com/ComputationalRobotics/XM-code
• https://arxiv.org/abs/2502.04640

Scenescript

• https://github.com/facebookresearch/scenescript
• Reconstructing Scenes With An Autoregressive Structured Language Model

Tightly Coupled Range Inertial Odometry and Mapping with Exact Point Cloud Downsampling

• https://staff.aist.go.jp/k.koide/assets/pdf/icra2025.pdf
• LiDAR-IMU SLAM framework with point cloud downsampling based on exact coreset extraction -- in short, a CPU version of GLIM that can even run on an affordable mini PC like one with an Intel N150!

Google Shopping's 3D generation

• https://research.google/blog/bringing-3d-shoppable-products-online-with-generative-ai/

• 1st Gen: NeRF

• 2nd Gen : NeRF + Diffusion

• 3rd Gen (Now): Video generation (Veo)

Maya4_ThirdGen.mp4

[U-AMC]

Interesting papers

PyRoki ("Python Robot Kinematics"): easier IK, trajectory optimization, motion retargeting... with an open-source toolkit on both CPU and GPU

JaArXmoRoP-ul3Uu.mp4

VINS-Multi: A Robust Asynchronous Multi-camera-IMU State Estimator

https://arxiv.org/pdf/2405.14539

FOCI: Trajectory Optimization on Gaussian Splats
https://rffr.leggedrobotics.com/works/foci/

[chae2chae]

[ICRA 2025] ForceMimic: Force-Centric Imitation Learning
with Force-Motion Capture System for Contact-Rich Manipulation