Twitter Beğeni Hilesi

Super excited to finally release our work, ReKep, a unified task representation using relational keypoint constraints. 🤖 rekep-robot.github.io A few key takeaways: 1. Building on the success of VoxPoser, VLM-generated code has proven to be extremely versatile in task

It’s very inspiring to see how general the formulation of constraint is using keypoint representation. Really like the pouring example, as it shows that keypoint-based constraints can handle orientation-related requirements. Check out Wenlong Huang ’s thread to see how it works!

Cool work from Wenlong! As emphasized in my ICLR2024 paper "Can Transformers Capture Spatial Relations between Objects?", spatial relation is a critical concept in robotics perception and planning modules!

Love the idea of integrating LVM & VLM with optimization! In the era of foundation models, optimization-based methods still play a role in robotics. Congrats to the authors!

Nice work! The clever and straightforward idea effectively leverages VLMs, enabling structured tasks to achieve zero-shot performance.

Very similar to CoPa (Constraints of Parts, cited as [96] in ReKep) by Haoxu Huang and Fanqi Lin CoPa formulate constraints in terms of vectors and points. ReKep use points only, but explore more complex functions in constraints. See: x.com/gao_young/stat…

Two robot arms move at the same speed, driven by different actuators with the same mass. The first arm collides with a table with a gentle tap. The second arm collides with the table, destroying both arm and table. Read this blog post to see why! 🦾💥 evjang.com/2024/08/31/mot…

Our lab’s new work that shows deeper integration of vision and robotic learning! 🤩🦾

Data is powerful but expensive for robotics. Continuing our VoxPoser from last year, by pushing what "zero-shot" methods can achieve and importantly, fall short, this sheds light on the complementary synergies b/t foundation models and large-scale data collection in robotics

Really nice article from @ericjang about actuators in robotics. Great to see 1X place such an emphasis on design. Safety through compliance and backdrivability were some of the core themes behind our BLUE project back in the day (berkeleyopenrobotics.github.io). Some thoughts 👇

Training on standard HOI datasets and executing robots without any fine-tuning ? We introduce General Flow, a 3D flow representation extracted directly from human datasets to achieve zero-shot human-to-robot skill transfer. Accepted by #CoRL2024. general-flow.github.io

Enjoying the fruits of my labor now that I finished my 12-legged 'Carpentopod' table project. See decarpentier.nl/carpentopod for project info, or see it live at Maker Days Eindhoven this 13 & 14 Sept.

I will be joining Engineering Dept as an Assistant Professor in spring 2025, together with Ayush Tewari. Clearly have been missing the good old UK rain after a wonderful year in California. Looking forward to opening this new chapter with brilliant colleagues and students!

Come check out o1-mini: SoTA math reasoning in a small package openai.com/index/openai-o… with Hongyu Ren Shengjia Zhao Eric Wallace & the rest of the OpenAI team

OpenAI Strawberry (o1) is out! We are finally seeing the paradigm of inference-time scaling popularized and deployed in production. As Sutton said in the Bitter Lesson, there're only 2 techniques that scale indefinitely with compute: learning & search. It's time to shift focus to

Today, I’m excited to share with you all the fruit of our effort at OpenAI to create AI models capable of truly general reasoning: OpenAI's new o1 model series! (aka 🍓) Let me explain 🧵 1/

What is a really really hard problem to work on in #AI? My own answer is Spatial Intelligence - a technology that could empower and enable countless possible use cases in creation, design, learning, AR/VR, robotics, and beyond. It’s a real honor that my cofounders Justin Johnson

🔥Spatial intelligence needs fast, *interactive* 3D world generation 🎮 — introducing WonderWorld: generating 3D scenes interactively following your movement and content requests, and see them in <10 seconds! 🧵1/6 Web: kovenyu.com/WonderWorld/ arXiv: arxiv.org/pdf/2406.09394