@boce_hu ISP can do the demos shown below with only approx 50 demos. Pretty amazing!
English
Robert Platt
74 posts
Robert Platt
@RobotPlatt
Computer Science Professor at Northeastern. Robotic manipulation, learning, perception, planning, control, and other stuff. My group: @HelpingHandsLab
Katılım Nisan 2022
157 Takip Edilen746 Takipçiler
Closed-loop visuomotor control with wrist cameras is widely adopted and powerful, but leveraging full 3D symmetry from only RGB is still hard.
Introducing our #NeurIPS2025 Spotlight paper, Image-to-Sphere Policy (ISP), for equivariant policy learning from eye-in-hand RGB images.
English
This paper was a finalist for the CoRL 2024 Outstanding Paper Award! Congrats everyone!
English
I'm excited about this work. The equivariant model significantly reduces the number of demos needed to train good diffusion policies.
Dian Wang@Dian_Wang_
Introducing Equivariant Diffusion Policy, a novel sample efficient BC algorithm based on equivariant diffusion. Our method leverages the symmetry in policy denoising to boost learning — needing 5x less training data in sim and mastering complex tasks in real-world with <60 demos.
English
Not only is this an interesting approach, but it works great! Outperforms everything we benchmarked against.
Dian Wang@Dian_Wang_
Check out our new grasping work, OrbitGrasp! It achieved a 98% grasping success rate by learning the SE(3)-equivariant grasp quality distribution using spherical harmonics. I’ll be at #CoRL2024 to present this paper.
English
Instead of inferring a desired object pose directly, this method "imagines" a reconstruction of the entire scene in the target pose. Surprisingly, we find that this improves sample efficiency, even though we are inferring more information.
Dian Wang@Dian_Wang_
Checkout our new work, Imagination Policy. We leverage a point cloud diffusion model to “imagine” a target scene, then use SVD to calculate rigid transformations that bring objects to the imagined scene as robot actions. More importantly, Imagination Policy is bi-equivariant!
English
Robert Platt retweetledi
#ICLR24
We proposed FourTran, a very sample-efficient 3D manipulation pick-place model.
1. It can learn a nontrivial 3D policy with less than 10 demos.
2. It represents 3D action distribution in Fourier Space.
Check it in the Poster Session 4 at 4:30 PM Vienna time (10:30 EDT)
English
Robert Platt retweetledi
Transporter network is great for robotic pick-place, but how to use it in 3D where rotating the kernel for entire SO(3) is infeasible? Our work performs cross correlation in Fourier space to facilitate reasoning over SO(3). Check out at #ICLR2024, @RobinSFWalters is presenting!
Haojie Huang@HaojieHuang13
#ICLR24 We proposed FourTran, a very sample-efficient 3D manipulation pick-place model. 1. It can learn a nontrivial 3D policy with less than 10 demos. 2. It represents 3D action distribution in Fourier Space. Check it in the Poster Session 4 at 4:30 PM Vienna time (10:30 EDT)
English
@HaojieHuang13 In short, I believe this is the best method out there for finding novel object grasps in point clouds. And the code is available: github.com/HaojHuang/Edge…
English
Excited to present our work - Edge Grasp Net: A Graph-Based SE(3)-invariant Approach to Grasp Detection at #ICRA2023!
A new grasping method first incorporates SE(3) invariance and is purely based on geometric features.
Project Website: haojhuang.github.io/edge_grasp_pag…
English
@HaojieHuang13 As far as I know, this method has the highest reported grasp success rates for the Berkley Adversarial objects for single view point clouds.
English
@SaulBadman12 @HelpingHandsLab I'm excited about this work. For the first time, we're using symmetric neural models to improve sample efficiency in behavior cloning of closed loop policies
English
Very excited to present our work at #ICRA2023! Our work introduces a novel 3D data augmentation method and equivariant policy learning, leading to high data efficiency. Our video shows how our method learns manipulation tasks using less than 10 demos, trained from scratch! 1/n
English
@DianWang1007 @HelpingHandsLab @andyzengtweets @HughWang19 @haosu_twitr @tesssmidt @erikjbekkers Looking forward to our workshop!
English
Excited to announce the #RSS2023 Workshop on Symmetries in Robot Learning! Submit your work and join us to explore how symmetries can help in robotics with our incredible speakers @andyzengtweets @HughWang19 @haosu_twitr @tesssmidt @erikjbekkers. Website: sites.google.com/view/rss23-sym.
English
@_davidklee @HelpingHandsLab It's easy to think that you can't leverage SO(3) symmetries when you're reasoning about 2D images. Not true! David got great results doing exactly that!
English
How can we efficiently model uncertainty over 3D object orientation given a single image? Our #ICLR23 work shows how to leverage SO(3) symmetry to better learn complex rotation distributions. My talk is on Tues. (May 2nd) in Oral3 Track2 (AD12), & Poster Session 3
English
@DianWang1007 @HelpingHandsLab Folks sometimes think that you can't apply symmetric neural models in real world scenarios bc the symmetry assumptions are not always perfectly met. This paper shows that it often works really well anyway even when there is significant mismatch between model and domain.
English
Facing an equivariant task with mismatched image & object transformations? Our #ICLR2023 paper discovers the power of equivariant network when enforcing symmetry to out-of-distribution data. Join my talk tmr (May 1st) in Oral2 Track4 (AD10) at 3:10 or drop by Poster Session 2! 1/
English
Robert Platt retweetledi
Registration is now open for Boston Symmetry Day on April 7th! We have an exciting lineup of speakers (@mario1geiger, @tesssmidt S Gortler, E Grigsby, @RobotPlatt, @_onionesque) + more to come.
Sign up here: forms.gle/8gZUh1BRL2FFat…
English
Robert Platt retweetledi
Hello world! The Boston Symmetry Group (pun intended) organizes events to foster community amongst Boston-area researchers and practitioners working on symmetry + machine learning.
Organizers: @StefanieJegelka @HLawrenceCS @dereklim_lzh @park_jungy @RobinSFWalters
English
Robert Platt retweetledi
Congrats to my PhD student Sammie Katt for passing his defense titled "Bayesian Partial Observable Reinforcement Learning." Very cool work, Sammie! (with @faoliehoek, @RobotPlatt, and Lawson Wong).
English
@GuanangS @HelpingHandsLab @SaulBadman12 @KhouryCollege @DianWang1007 @_davidklee @XupengZ @RobinSFWalters Great job, Mingxi!
Suomi
Helping Hand member @SaulBadman12 talks about his approach for Equivariant Imitation Learning at the @KhouryCollege Research Apprenticeship Showcase. Work with @DianWang1007 @GuanangS @_davidklee @XupengZ @RobinSFWalters @RobotPlatt
English
Robert Platt retweetledi
Interested in partial observability? Come to my online poster session at @CoRL2022 (online session 2, Sunday, Dec 18, 7:30-8.00 AM, Auckland time, i.e., Saturday, Dec 17, 1:30 - 2:00 PM, EST time) w/ @RobotPlatt @cjdamato @DianWang1007 @AndreaBaisero @LLPR_NEU @HelpingHandsLab.
Hai Nguyen@HaiNguy69482974
Glad to share our #CoRL22 paper arxiv.org/abs/2211.01991 about using MDP solutions to efficiently learn POMDPs during offline training. We proposed a SAC-like agent that balances between acting like an MDP expert and for environment rewards.
English
@DianWang1007 This CoRL paper is really cool. We use symmetric neural models to improve sample efficiency of RL so that we can learn directly on physical robots. It's even cooler in person b/c you can see it learn in front of your eyes!
English
How can we do real-world RL where acquiring samples take a lot of time? Our #CoRL2022 paper shows that equivariant method improves sample efficiency and facilitates real-world RL. I will present this work in poster session 1 tmrw at the poster lobby (paper 246), have a look! 1/
English
@HaiNguy69482974 Hai is presenting this work this week at CoRL. He's using solutions to a fully observable version of the problem to solve a POMDP.
English
Glad to share our #CoRL22 paper arxiv.org/abs/2211.01991 about using MDP solutions to efficiently learn POMDPs during offline training. We proposed a SAC-like agent that balances between acting like an MDP expert and for environment rewards.
English