Liqian Ma

Robotics PhD student @ Georgia Tech. BEng @ Tsinghua

[Liqian Ma @LiqianMa4]

(8/8) Joint work with @saxenavaibhav11, @utkarshm0410, and @ShuoCheng94, advised by @YongxinChen1 and @danfei_xu.

📄 https://t.co/vnkQympgWx
🌐 https://t.co/ZLy0PeJ8Ln

Poster at #CoRL2026: Nov 11, Poster session 3

[Liqian Ma @LiqianMa4]

In sim-and-real co-training, robots learn from a few real demos plus ample sim data. But which sim data should we generate?

MetaGenSim meta-learns what to simulate: 2.5× the real-world success of uniform sim data, matching expert hand-tuning.

https://t.co/ZLy0PeJ8Ln
#CoRL2026

[Liqian Ma @LiqianMa4]

(7/8) MetaGenSim can still generate the closest task data even when the real task has no exact sim counterpart.

For the novel real task mugTray, it generates mostly mugBin (grasp the mug) and bowlTray (place on a tray) data, beating uniform generation.

[Shuo Cheng ✈️CoRL2026 @ShuoCheng94]

Can large-scale sim data enable real-world generalization?🤔
In our new work, we introduce a generalizable domain adaptation setting, where policies must handle real-world situations never presented in the real training data. (1/n)