Wenxuan Zhou @Wenxuan_Zhou Twitter profile | Pikagi

Pikagi

Wenxuan Zhou

@Wenxuan_Zhou

1,749

Followers

265

Following

14

Media

55

Statuses

Research Scientist at @Meta GenAI; Ph.D. in Robotics @CarnegieMellon ; Former Research intern at Google @DeepMind

https://t.co/QkKbrJlja1

Joined October 2015

Don't wanna be here? Send us removal request.

Pinned Tweet

@Wenxuan_Zhou

Wenxuan Zhou

9 months

Life updates: Successfully finished my Ph.D. thesis defense! It’s been an incredible journey of exploring the possibilities of robots and RL. I’m actively seeking full-time scientist/engineer positions in AI/Robotics. Looking forward to new adventures! ⛵️

Tweet media one

48

24

722

Last Seen Profiles

@BeerhillGangTV

@DailyNadeko

@yellaa88

@MayankRMFC

@hwuthi16087187

@turbanlisever69

@jandakembangstw

@tjgnan

@oyoyoxxx

@Footie_Xpress

@polsek_muntok

@ShamsaldeenM

@andrew_milito

@SeyedamiHoseyni

@Evan_LRodriguez

@Inlonte

@ddevineee

@FinesttCartier

@pvs_oppo98397

@momofifs

@Zenomorph255112

@BobaPhil

@pnwnd

@pengen_stw

@CuntressLuann

@pG6gEmay99t87C

@maxrwag

@tscrystalskies

@Stella80931044

@ZeEshanM01

@AnnePhilli17754

@____Pingu____

@stormaxou

@monkey_ZmLN

@theyloveebrit

@luver_ra

@Wenxuan_Zhou

Wenxuan Zhou

1 year

How can robots learn generalizable manipulation skills for diverse objects? Going beyond pick-and-place, our recent work “HACMan” enables complex interactions for unseen objects, such as flipping, pushing, or tilting, using spatial action maps + RL with point clouds. (w/ @MetaAI )

7

57

317

@Wenxuan_Zhou

Wenxuan Zhou

2 years

Are simple grippers limited to simple motions such as pick-and-place? Our work to be presented at #CoRL2022 demonstrates that RL can be used to enable a parallel gripper to find interesting strategies to exploit the environment to enhance its “dexterity”. A thread:

2

35

205

@Wenxuan_Zhou

Wenxuan Zhou

9 months

Code released!

Tweet card media

GitHub - HACMan-2023/HACMan: HACMan: Learning Hybrid Actor-Critic Maps for 6D Non-Prehensile...

HACMan: Learning Hybrid Actor-Critic Maps for 6D Non-Prehensile Manipulation - HACMan-2023/HACMan

@Wenxuan_Zhou

Wenxuan Zhou

1 year

How can robots learn generalizable manipulation skills for diverse objects? Going beyond pick-and-place, our recent work “HACMan” enables complex interactions for unseen objects, such as flipping, pushing, or tilting, using spatial action maps + RL with point clouds. (w/ @MetaAI )

7

57

317

0

10

89

@Wenxuan_Zhou

Wenxuan Zhou

2 years

Additional problems arise in lifelong learning with non-stationary dynamics if we use off-policy data. Check out our work "Forgetting and Imbalance in Robot Lifelong Learning with Off-policy Data" at @CoLLAs_Conf , from my internship at @DeepMind !

Tweet media one

2

9

67

@Wenxuan_Zhou

Wenxuan Zhou

3 years

Glad to hear that! Harshit will present our work on Wednesday during Oral Session 5. And please join our poster session right after that!

@corl_conf

Conference on Robot Learning

3 years

Congratulations to #CoRL2021 best paper finalist, "Learning Off-Policy with Online Planning", Harshit Sikchi, Wenxuan Zhou, David Held. #robotics #learning #award #research

Tweet media one

7

25

138

0

3

47

@Wenxuan_Zhou

Wenxuan Zhou

4 years

Excited to be presenting our work on offline RL at the plenary session at #CoRL2020 @corl_conf ! We propose to learn a Policy in the Latent Action Space (PLAS) to naturally avoid out-of-distribution actions. w/ Sujay Bajracharya, @davheld

Tweet media one

1

6

23

@Wenxuan_Zhou

Wenxuan Zhou

2 years

Virtually attending CoRL to present this work: - Oral session 6 - Manipulation and Grasping - Poster session 6 (OGG (260), Room 038) - Online poster session 1 on Pheedloop at 9:35 PM EST on Friday Happy to chat about our work, anything robot learning 🤖, or just to say hi👋

@Wenxuan_Zhou

Wenxuan Zhou

2 years

Are simple grippers limited to simple motions such as pick-and-place? Our work to be presented at #CoRL2022 demonstrates that RL can be used to enable a parallel gripper to find interesting strategies to exploit the environment to enhance its “dexterity”. A thread:

2

35

205

0

0

22

@Wenxuan_Zhou

Wenxuan Zhou

2 years

Join us at the Offline RL workshop tomorrow! Check out the schedule here:

@OfflineRL

OFFLINE RL NeurIPS Workshop

2 years

We are back for 3rd Offline RL workshop, which be held in-person at @NeurIPSConf 2022. The theme of the workshop is using "Offline RL as a Launchpad". Call for papers and details at .

1

5

33

0

2

20

@Wenxuan_Zhou

Wenxuan Zhou

1 year

Check out the paper and the website for more information and video results showing HACMan generalizing to different objects and goals! w/ @bwww08 , Fan Yang, @chris_j_paxton , @davheld

0

1

13

@Wenxuan_Zhou

Wenxuan Zhou

1 year

Our object-centric action representation has two benefits. It is… 1. Spatially-grounded: because the learned contact location is selected from the observed object points. 2. Temporally-abstracted: because we focus only on learning the contact-rich portions of the action.

Tweet media one

1

1

9

@Wenxuan_Zhou

Wenxuan Zhou

2 years

@CoLLAs_Conf @DeepMind Big thanks to all my collaborators: @sbohez , Jan Humplik, Abbas Abdolmaleki, @drao64 , @markus_with_k , Tuomas Haarnoja, Nicolas Heess. It was an awesome internship experience!

0

0

5

@Wenxuan_Zhou

Wenxuan Zhou

1 year

It looks like the challenge of manipulation in the video is mainly to challenge the locomotion skill. How far are we from the wow moment of skillful manipulation as mind-blowing as atlas backflipping? 🤔

@Vikashplus

Vikash Kumar

1 year

🔍Perspective on latest Atlas results - Good old SysID+Sim2Real - Estimation+Control - Model based TrajOpt, not MBRL - Predictive control on prescribed (not learned) model - Composition using known skills - Generalize skills, not tasks 🤔 What did I miss?

3

1

17

0

0

6

@Wenxuan_Zhou

Wenxuan Zhou

2 years

Check out the paper “Learning to Grasp the Ungraspable with Emergent Extrinsic Dexterity” to be presented at #CoRL2022 ! (w/ @davheld )

Learning to Grasp the Ungraspable with Emergent Extrinsic Dexterity

[Press Coverage] IEEE Spectrum: Robots Grip Better When They Grip Smarter

sites.google.com

0

0

5

@Wenxuan_Zhou

Wenxuan Zhou

2 years

Takeaways: 1. Simple grippers are underrated when combined with a skillful policy. 2. Robots should exploit extrinsic dexterity for more skillful manipulation. 3. RL can generate policies with emergent extrinsic dexterity that transfer to the real world.

1

2

4

@Wenxuan_Zhou

Wenxuan Zhou

1 year

We find that defining the right action space is crucial for learning a manipulation task. We explore an object-centric action representation in RL that consists of selecting a contact location on the object and a set of parameters describing the robot's movement after contact.

Tweet media one

1

0

4

@Wenxuan_Zhou

Wenxuan Zhou

2 years

We develop a system with RL to show “extrinsic dexterity”, e.g. taking advantage of the extrinsic environment. In the “Occluded Grasping” task, the policy learns to pushing the object against the wall to rotate and then grasp it without any reward terms on extrinsic dexterity.

Tweet media one

1

0

4

@Wenxuan_Zhou

Wenxuan Zhou

1 year

@arnav_wadhwa_ @MetaAI Multi-fingered hands may allow a wider variety of motions and have more tolerance (picking an object with a multi-fingered hand can be less sensitive to object shapes than a simple gripper). However, they are more expensive, easier to break, and have a bigger sim2real gap.

1

0

2

@Wenxuan_Zhou

Wenxuan Zhou

1 year

@arnav_wadhwa_ @MetaAI Nonetheless, both types of hands are definitely worth exploring!

0

0

2

@Wenxuan_Zhou

Wenxuan Zhou

1 year

@arnav_wadhwa_ @MetaAI So IMO, we need good reasons to pay the cost of maintaining the multi-fingered hands. If a task can be done with simple hands, then we should go with simple hands! My research has been trying to push this limit of what a simple hand can achieve.

1

0

2

@Wenxuan_Zhou

Wenxuan Zhou

1 year

HACMan outperforms the baselines, with a larger margin for more challenging tasks. Success rates for simple tasks - pushing a single object to an in-plane goal - are high for all methods, but only HACMan achieves high success rates for 6D alignment of diverse objects.

Tweet media one

1

1

2

@Wenxuan_Zhou

Wenxuan Zhou

9 months

@QinYuzhe thank you!!

0

0

0

@Wenxuan_Zhou

Wenxuan Zhou

4 years

arXiv:

0

0

1

@Wenxuan_Zhou

Wenxuan Zhou

4 years

Plenary Talk: Nov 18th (Wed), 10:45 - 11:00 AM EST Interactive Session: Nov 18th (Wed), 2:10 - 2:40 PM EST

1

0

2

@Wenxuan_Zhou

Wenxuan Zhou

2 years

Website: Virtual poster session: 4 AM - 7 AM EST, Aug 19th 2022 Feel free to email me (wenxuanz @andrew .cmu.edu) if you'd like to chat but cannot make it to the poster session!

0

0

1

@Wenxuan_Zhou

Wenxuan Zhou

1 year

@SusieMorga37205 @MetaAI Thank you!

0

0

1

@Wenxuan_Zhou

Wenxuan Zhou

1 year

@shi40553510 @MetaAI Thank you! We expect to release the code on the project website within the next 1-2 weeks. Please stay tuned!

1

0

1

@Wenxuan_Zhou

Wenxuan Zhou

1 year

@devesh_dkj @davheld Thanks for sharing! It looks pretty related.

0

0

1

@Wenxuan_Zhou

Wenxuan Zhou

1 year

With off-policy RL, given a point cloud, the actor outputs per-point motion parameters (Actor Map) while the critic outputs per-point Q-values (Critic Map). The Critic Map is not only used to update the actor but also serves as the scores for selecting the contact location.

Tweet media one

1

1

1

@Wenxuan_Zhou

Wenxuan Zhou

2 years

@tonyzzhao @Stanford That's great to hear. Looking forward to it!

0

0

1

@Wenxuan_Zhou

Wenxuan Zhou

2 years

The policy trained in simulation is zero-shot transferred to a physical robot. It demonstrates dynamic and contact-rich motions with a simple gripper that generalizes across objects with various size, density, surface friction, and shape with a 78% success rate.

1

0

1

@Wenxuan_Zhou

Wenxuan Zhou

1 year

We evaluate our method with a 6D object pose alignment task with randomized initial poses, randomized 6D goals, and diverse unseen objects in both simulation and in the real world.

Tweet media one

1

1

1