Yu Xiang

@YuXiang_IRVL

I do research in robotics and computer vision @IRVLUTD, Ex- research scientist @NVIDIA, postdoc @uwcse and @StanfordAILab, phd @UMich, all views my own.

Dallas, TX

yuxng.github.io

Joined October 2023

1,141 Following

7,519 Followers

Replies Reposts Media

Pinned
Yu Xiang
@YuXiang_IRVL
Jun 25
Some of my recent thoughts on building foundation models for robot manipulation: • Don't commit the model to a single robot hardware. Hardware will continue to evolve, and models should transfer across platforms. • Don't commit to two-finger grippers. Dexterous hands are the
7
Yu Xiang
@YuXiang_IRVL
Oct 20
I feel Agentic AI and Robotics will eventually converge. Both are about reasoning and acting, just in different environments. Acting in the physical world is much more challenging, involving embodiment, perception, control, safety, and physical interactions.
9
Yu Xiang
@YuXiang_IRVL
Oct 17
All set for CoRL. Thanks to @lfcasas7 and @ZQ_fantastic for covering my lectures while I’m away. Looking forward to meeting everyone at CoRL!
2
Yu Xiang
@YuXiang_IRVL
Oct 16
This work gives an open-weight model 20 tools for robot manipulation. One trend I see is designing tools and using LLMs as reasoners. Another approach is to provide just one tool: "move hand(s)." The LLM then acts more like a VLA model. I personally prefer the second approach.
Hao AI Lab
@haoailab
Oct 16
00:53
GPT and Claude now operate real robot arms. What about open-weight models? 🤔 Introducing ORA: Open Robot-Use Agents. An open harness powered by @deepseek_ai V4.1-Flash, ORA operates robots from visual input—zero-shot, no VLA—and competes with frontier-model agents.
3
Yu Xiang
@YuXiang_IRVL
Oct 15
Want to know how to use models like Astra for robot manipulation? Need to check this report below 3 approaches with ablations: 1. Direct robot actions 2. Codex with tools 3. Codex with a pretrained policy Option 2 performs best, due to Codex's persistent context and tool use.
Jenai Xuning Yang
@xuningy
Oct 14
Can and should a general-purpose model serve as a robot policy? In Astranomy, we study GPT-6 Astra for general pick-and-place manipulation. Astra sets a new SOTA on RoboLab-120, but real-time robot control remains an open challenge. Full tech report: research.nvidia.com/labs/srl/proje…
00:00
8