Wan𝕏in Jin
@jinwanxin
Followers
386
Following
231
Media
8
Statuses
62
Asst. Prof @ASU | Working on robotics | Prev. @GRASPlab | PI of Intelligent Robotics and Interactive Systems (IRIS) Lab
Tempe
Joined April 2014
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
lisa
• 562384 Tweets
Colombia
• 193966 Tweets
#السعوديه_البحرين
• 174307 Tweets
England
• 164351 Tweets
سالم
• 113694 Tweets
#VSFashionShow
• 96170 Tweets
Tuchel
• 94411 Tweets
مانشيني
• 59067 Tweets
ADRIANA LIMA
• 48584 Tweets
Amari Cooper
• 43655 Tweets
Tyla
• 40077 Tweets
#GHLímite6
• 39620 Tweets
Córdoba
• 38715 Tweets
GIGI
• 38529 Tweets
المنتخب السعودي
• 32426 Tweets
James Rodríguez
• 31709 Tweets
Durán
• 30942 Tweets
Cher
• 27925 Tweets
Tapia
• 27398 Tweets
#فزعه_الشعب_السعودي
• 23499 Tweets
EU TE TRATO MAL
• 20997 Tweets
Gareca
• 16266 Tweets
Lucho
• 14996 Tweets
Dávila
• 13909 Tweets
Barranquilla
• 13376 Tweets
Vidal
• 11203 Tweets
Pinned Tweet
🚀 Can
#MPC
rival and even surpass
#ReinforcementLearning
in solving
#DexterousManipulation
? Our answer is a resounding YES! PROUD to share: 🔥Complementarity-Free Multi-Contact Modeling and Optimization, our latest method that sets shattering benchmarks in various challenging
4
36
226
🚀Can a robotic hand master
#DexterousManipulation
in just 2 mins? YES! 🎉Excited to share our work “ContactSDF,” a physics-inspired representation using signed distance functions (
#SDFs
) for dexterous manipulation, from
#geometry
to
#MPC
. 🔥Watch the UNCUT video of Allegro Hand
1
10
57
Our work “Safe Pontryagin Differentiable Programming” () is accepted by
#NeurIPS2021
. With Prof. George J. Pappas
@pappasg69
& Prof. Shaoshuai Mou, we provide a safe differentiable framework to solve a board class of safety-critical learning & control tasks
1
3
20
Robots can infer rewards from human feedback, but what about safety boundaries? Our work, Safe MPC Alignment (), led by
@ZhixianXie_ASU
, collab’d with
@asuriselab
, Yi Ren,
@zhaoran_wang
, and
@pappasg69
, shows it’s possible and can be human-data efficient!
1
4
12
What a great experience to work with Michael
@MichaelAPosa
for efficient learning & control of dexterous manipulation! Contact-rich manipulation is computationally tough due to its hybrid nature. But our idea is simple: far fewer hybrid modes could be enough to achieve many tasks
I'm excited to share our latest preprint. We use 4 minutes of experiential data to learn a model for robust real-time manipulation of a previously unknown object. Work led by
@jinwanxin
, supported by
@ToyotaResearch
. 1/
1
9
72
0
0
11
Can a robot learn its cost function from a human user's DIRECTIONAL correction? Our new T-RO paper (with
@todd_murphey
, Z. Lu, S. Mou) provides a simple, efficient, but comprehensive solution to this question. Below is the intro video. 1/5
2
2
9
The key is our novel "Complementarity-Free Multi-Contact model", which effectively models contact-rich interactions (particularly handling friction), while being very simple and optimizer-friendly!
🔗 Dive deeper:
• Preprint:
• Code:
0
0
6
Our method enables robots to learn safety constraints with a small handful of user corrections. It is certifiable, providing an upper bound on the number of human corrections needed for successful learning or indicating if the true constraint is outside the hypothesis space.
1
0
3
@Tesla
My model 3 failed on installing update 2023.44 30.... multiple times. I want to know why....
0
0
2
Our algorithm is simple to implement and intuitive: repeatedly “cut” through the space of possible objective parameters in the direction given by the human corrections until the remaining space is small enough. 3/5
1
0
2
@ChongZitaZhang
Oops, the cat’s out of the bag! 🎉 Haven’t even tweeted it yet, "secrets" slipped out. LOL
0
0
2
The user study and real-world experiments indicate our method is significantly more effective (higher success rate), efficient/effortless (less human corrections needed), and potentially more accessible (fewer early wasted trials) than the state-of-the-art. 4/5
1
0
2
🔗 Dive deeper:
Web:
Preprint:
Code:
YouTube:
0
0
2
5/5
IEEE Xplore:
Arixv:
Code:
Video:
0
0
2
Our method learns robot cost function from human correction using only the directionality of correction, not the magnitude. We show learning from the directionality of correction has provable convergence, whereas prior works using the magnitude do not and can be inefficient. 2/5
1
0
2
1
0
1
@AlexBeatson
@zhaoran_wang
Hi Alex, (13) and (18) are for the general case where the parameterized objective function J(\theta) in (1) is not zero. For the OC, J(\theta)=0 and (13b-d) and (18) are trivialized. The auxiliary control system for OC is simplified to (21), which is forward. Plz see Algorithm 5
0
0
1
@AlexBeatson
@zhaoran_wang
Hi Alex, one additional comment from mine, as i previously thought you were only talking about the OC mode here. For the general PDP framework, the difference from the adjoint method is: PDP also differentiates the adjoint state as well. Please see Equ (13b).
0
0
1
@_sam_sinha_
@zhaoran_wang
Thank for your comments. The learned optimal policy is initial-condition dependent, since the learned policy is a result of a finite-horizon optimal control system. However, given different initial conditions, the proposed PDP can quickly learn the corresponding optimal policy.
1
0
1
@Bluengine
@zhaoran_wang
Thanks. Differentiating through the PMP can be readily extended to continuous-time systems: just differentiating through the continuous-time version of the PMP, and then the very similar results can be achieved. We will update soon to Arxiv.
0
0
1