phd @CarnegieMellon | Intern at @Amazon FAR| Previously @Tsinghua_Uni | Studying Control & Robot Learning
Pittsburgh
Joined May 2024
- Tweets43
- Following371
- Followers386
- Likes175
Pinned Tweet
Common belief: Robot policies need continuous action heads; discrete actions are ๐ฅ๐๐ฌ๐ฌ ๐ง๐๐ญ๐ข๐ฏ๐ ๐๐ง๐ ๐ฅ๐๐ฌ๐ฌ ๐ฉ๐ซ๐๐๐ข๐ฌ๐.
โWe find policies w/ the simplest discrete representations ๐ฆ๐๐ญ๐๐ก ๐จ๐ซ ๐๐๐๐ญ continuous onesโin sim & real, in single-step & diffusion/flow settings.
Here's why ๐งต
zeji yi retweeted
@YananSuiโs group built this.
zeji yi retweeted
With @JayHe748646 @KehanWen170077 @leggedrobotics , we push humanoid mobility to next frontier.
Achieved by a generalist contact planner + a generalist tracker + a generalized neural mapping, warmstarted from few human data.
Website: leggedrobotics.github.io/ACPโฆ
Promo Video ๐
11/12 Takeaway: we may need to ๐ซ๐๐ญ๐ก๐ข๐ง๐ค ๐ฐ๐ก๐๐ญ๐ก๐๐ซ ๐ซ๐จ๐๐จ๐ญ ๐๐๐ญ๐ข๐จ๐ง๐ฌ ๐ฆ๐ฎ๐ฌ๐ญ ๐๐ ๐๐จ๐ง๐ญ๐ข๐ง๐ฎ๐จ๐ฎ๐ฌ.
Uniform bins, the simplest discrete actions, match or beat continuous heads once their tokens are kept consistent, through guidance or iterative refinement.
Perhaps robot actions are ๐ง๐จ๐ญ ๐ญ๐ก๐๐ญ ๐๐ข๐๐๐๐ซ๐๐ง๐ญ ๐๐ซ๐จ๐ฆ ๐ฅ๐๐ง๐ ๐ฎ๐๐ ๐ ๐๐ง๐ ๐ฏ๐ข๐ฌ๐ข๐จ๐ง: they can be tokens too.
ALT Robot actions may not need to be continuous
12/12 Paper: arxiv.org/abs/2610.20429
Project page, 5-min explainer and all rollout videos: iscoyizj.github.io/discrete-โฆ
Code and checkpoints: coming soon.
Joint work with @AstonishingWolf, Xinyang Zhang, @KoushilSreenath, @guannanqu, @peterxichen, @max_simchowitz, @GuanyaShi at CMU and Amazon FAR.
zeji yi retweeted
We're hiring interns (and full-times) all year long! Please email me if interested.
Also happy to talk about some of our other latest work like DIAL-MPC: a Diffusion-style annealing + sampling-based MPC lecar-lab.github.io/dial-mpcโฆ.
I'll be in Vancouver for #NeurIPS2024 presenting Model-Based Diffusion (MBD) lecar-lab.github.io/mbd/ a diffusion-based training-free trajectory optimization method. I'm eager to chat about the latest in control and learning theory - please DM me if interested!
Also, I'm actively seeking a #Summer2025 internship in these areas. Catch me at the poster session on Wed, 11 Dec, Poster Session 1 West.
I'll be in Vancouver for #NeurIPS2024 presenting Model-Based Diffusion (MBD) lecar-lab.github.io/mbd/ a diffusion-based training-free trajectory optimization method. I'm eager to chat about the latest in control and learning theory - please DM me if interested!
I'll be in Vancouver for #NeurIPS2024 presenting Model-Based Diffusion (MBD) lecar-lab.github.io/mbd/ a diffusion-based training-free trajectory optimization method. I'm eager to chat about the latest in control and learning theory - please DM me if interested!
zeji yi retweeted
Why do we treat train and test times so differently?
Why is one โtrainingโ and the other โin-context learningโ?
Just take a few gradients during test-time โ a simple way to increase test time compute โ and get a SoTA in ARC public validation set 61%=avg. human score! @arcprize
zeji yi retweeted
๐จ๐จ One Policy, Many Modes!
Check our new versatile humanoid wholebody controller๐
๐ค Humanoids are expected to be the ultimate embodiment for general-purpose robots, but reaching that potential means adapting to diverse tasksโnavigation, tabletop manipulation, loco-manipulationโeach needing unique control modes.
๐ค Can one controller handle it all?
โจ Introducing HOVER:
๐งฉ Unified Command Space: Supports kinematic position, joint angle, and root tracking in a unified versatile policy.
๐ช Generalist Controller: HOVER outperforms previous specialized policies trained on specific modes (ExBody, H2O, OmniH2O, HumanPlus).
๐ Website: hover-versatile-humanoid.gitโฆ