Boyuan An @AnthonyInROS2

boyuan-an.github.io Joined December 2024

Tweets

10
Followers

14
Following

128
Likes

26

Boyuan An @AnthonyInROS2

5 days ago

@LinzhanMou Incredible work!

1 0 1 977 0

View Details

🔥 Can we scale tactile learning without collecting every possible contact in the real world? Introducing ControlTac — a tactile data generation framework that expands a single reference tactile image into diverse observations with controllable 3D contact forces and contact

1 2 8 330 3

1 0 0 70 0

View Details

Boyuan An @AnthonyInROS2

3 weeks ago

@JunxiaoLin_zju @Modern_Gangster @SharpaRobotics Great work, junxiao!(ฅ⁍̴̀◊⁍̴́)و ̑̑

0 0 0 13 0

View Details

Boyuan An @AnthonyInROS2

3 weeks ago

Really glad to be part of this work with @Modern_Gangster and the team!

W. Tianyue @Modern_Gangster

3 weeks ago

Can generated videos help us train more general dexterous controllers? 1/3: GALATEA - a first step toward a general-purpose dexterous tracker. Generated videos play two roles: they supply diverse motion references during training, and propose motion plans at deployment. We

1 1 12 3K 2

0 0 2 30 0

View Details

Boyuan An @AnthonyInROS2

3 weeks ago

@Modern_Gangster Really glad to be involved in this work and to take a small step together toward more general dexterous manipulation. Excited to see where this direction goes!

0 0 1 59 0

View Details

Boyuan An @AnthonyInROS2

3 weeks ago

Thanks for sharing our work! This is an initial exploration of whether a General Dexterous Tracker can enable more generalizable dexterous manipulation across different tasks and objects. We hope this can be a step toward more scalable and versatile robot manipulation.

Techniahqrobot | humanoid robots @techniahqrobot

3 weeks ago

3 5 12 7K 4

0 0 1 45 0

View Details

Techniahqrobot | humanoid robots @techniahqrobot

3 weeks ago

A generated video showed a robot how to grasp a jar. The interesting part came after the video. GALATEA uses Seedance 2.0 to generate manipulation clips, reconstructs the 3D motion of the hand and object, then tries to make that same interaction physically work inside PhysX before sending it to a real robot. They generated around 2,500 clips. About 2,000 produced usable hand-object trajectories, and more than 1,500 could actually be grounded in simulation. For the real-world tests, they used a Franka Research 3 with a 22-DoF Sharpa Wave Hand. One detail I really like is that GALATEA keeps the finger motion. So the reference contains more than "move this object from A to B." It can preserve how the object is held. Grab the container around the body. Hold it by the neck. Approach it with a particular finger configuration. Those choices matter a lot once you leave video generation and start dealing with contact, friction and joint limits. On 40 unseen real-world plans, they report 27 successful executions. That makes generative video much more interesting for robotics. You are no longer looking at a clip only as pixels. You can start treating it as a rough motion specification that still has to survive physics. @AnthonyInROS2 @YiMaTweets @kaiwynd

3 5 12 7K 4

View Details

Boyuan An @AnthonyInROS2

7 months ago

@shieldkeke Cool work on whole-body loco-manipulation!

0 0 0 17 0

View Details

Yeke Chen @shieldkeke

7 months ago

We're thrilled to present our work on a humanoid robot mastering badminton skills! 🏸🤖 arxiv: arxiv.org/abs/2602.08370 page: astrorix.github.io/LHBS/

1 1 4 204 0

View Details

Boyuan An @AnthonyInROS2

7 months ago

@OWW Thanks for sharing our work! Please feel free to visit our project website ai4ce.github.io/EgoPush/ for more details. We are currently preparing the code release. Should you have any questions, please do not hesitate to reach out to us via email.