Lyuxing He

[Lyuxing He @LyuxingHe]

@WangLingkan Thx!

[Lyuxing He @LyuxingHe]

How can robots place objects precisely when there are tight tolerances and changing object geometry, for example, reliably inserting a connector into a socket?

Introducing TAX-DPD, a hierarchical point-diffusion method for object-centric goal prediction. TAX-DPD achieves 80-100% accuracy on high precision insertion tasks on the NIST task board, and also only a 3% error rate on the RPDiff object placement benchmark that requires generalizing over object geometry.  #ICRA2026

Project: https://t.co/kBpLSBsJbn 
Paper: https://t.co/LprP4lF9s4 
Code: https://t.co/bPsCnoOzWN

[Lyuxing He @LyuxingHe]

To recap: TAX-DPD predicts object goal configurations with a two-stage hierarchy.

1️⃣ Stage 1 uses a scene-level Dense GMM to sample multimodal placement frames.

2️⃣ Stage 2 uses local disentangled point diffusion to refine object geometry and placement frame.

Together, this improves coverage, precision, and geometry generalization across rigid object placement tasks in RPDiff, real world industrial-level NIST insertions, and deformable cloth hanging tasks in DEDO.

🙏 Many thanks to the co-authors @eywcai, Shobhit Aggarwal, Jianjun Wang, and our advisor @davheld.

[Lyuxing He @LyuxingHe]

The point-cloud goal formulation also relaxes the rigid-object assumption.

On modified DEDO HangProcCloth-DH, TAX-DPD predicts a non-rigid cloth goal configuration, then a goal-conditioned DP3 policy executes the placement.

Compared with TAX3D, success improves from 50% to 78%, coverage RMSE drops from 0.87 to 0.50, and precision RMSE drops from 1.34 to 0.58.

[Kallol Saha @_ksaha]

🚨Introducing SPOT: Search over Point Cloud Object Transformations. SPOT is a combined learning-and-planning approach that searches in the space of object transformations.

Website: https://t.co/VwiyDI15FC
Paper: https://t.co/GmdF7hN0FG
Code: https://t.co/YC5sO9OZ0K

[Yishu Li @LisaYishu]

A closed door looks the same whether it pushes or pulls. Two identical-looking boxes might have different center of mass. How should robots act when a single visual observation isn't enough?

Introducing HAVE 🤖, our method that reasons about past interactions online! #CORL2025

[Divyam Goel @divyamgo10]

How do we discover a robot's failure modes before deploying it in the real world? Standard benchmarks often don't capture the full picture, leaving policies vulnerable to plausible variations in object shape.

Thrilled that our work, "Geometric Red-Teaming for Robotic Manipulation," has been accepted as an oral presentation at #CoRL2025! We introduce a framework to automatically find these geometric blindspots.
https://t.co/mh84bi7Wbp
🧵

[Alexis Hao @hao_alexis]

Introducing FMVP: a method that adapts to natural arm motions during robot-assisted dressing.

Pre-trained on vision in sim, fine-tuned with limited real-world vision+force data, and tested in a 12-user, 264-trial study, FMVP is robust across garments and motions. #CoRL2025