Soroush Nasiriany
@snasiriany
Followers
1K
Following
2K
Media
19
Statuses
159
PhD student @UTAustin. Building foundation models for robots.
Austin / Bay Area
Joined September 2014
Iโm excited to introduce RoboCasa, a large-scale simulation framework for everyday tasks. Scaling is the key driving force to unlocking generalist robots, and RoboCasa leverages simulation to take scaling to a whole new level. A short ๐งต
10
54
266
RT @stepjamUK: ๐'๐๐ฒ ๐ต๐ฒ๐ฎ๐ฟ๐ฑ ๐๐ต๐ถ๐ ๐ฎ ๐น๐ผ๐ ๐ฟ๐ฒ๐ฐ๐ฒ๐ป๐๐น๐: "๐ช๐ฒ ๐๐ฟ๐ฎ๐ถ๐ป๐ฒ๐ฑ ๐ผ๐๐ฟ ๐ฟ๐ผ๐ฏ๐ผ๐ ๐ผ๐ป ๐ผ๐ป๐ฒ ๐ผ๐ฏ๐ท๐ฒ๐ฐ๐ ๐ฎ๐ป๐ฑ ๐ถ๐ ๐ด๐ฒ๐ป๐ฒ๐ฟ๐ฎ๐น๐ถ๐๐ฒ๐ฑ ๐๐ผ ๐ฎ ๐ป๐ผ๐๐ฒ๐น ๐ผ๐ฏ๐ท๐ฒ๐ฐ๐ - ๐๐ต๐ฒ๐๐ฒ ๐ป๐ฒ๐ ๐ฉ๐๐ ๐บ๐ผ๐ฑโฆ.
0
53
0
RT @saxenavaibhav11: Large robot datasets are crucial for training ๐คfoundation models. Yet, we lack systematic understanding of what data mโฆ.
0
44
0
RT @huihan_liu: Meet Casper๐ป, a friendly robot sidekick who shadows your day, decodes your intents on the fly, and lends a hand while you sโฆ.
0
37
0
RT @hellorobotinc: ๐Stretch MuJoCo v0.5 is released! It's a high-fidelity simulation of Stretch 3. Hereโs what's new:. โข ROS2 and Python liโฆ.
0
17
0
RT @RoboPapers: Ep#11 with @snasiriany @SteveTod1998 @abhirammaddukur @Lawrence_Y_Chen on Sim-and-Real Co-Training: A Simple Recipe for Visโฆ.
0
9
0
Itโs not a matter of if, itโs a matter of when, video models and world models are going to be a central tool for building robot foundation models.
Introducing ๐๐ซ๐๐๐ฆ๐๐๐ง!. We got humanoid robots to perform totally new ๐ฃ๐๐๐๐ in new environments through video world models. We believe video world models will solve the data problem in robotics. Bringing the paradigm of scaling human hours to GPU hours. Quick ๐งต
0
1
11
Collecting real-world data is hard to scale. We show how co-training with large-scale simulation data from RoboCasa can significantly boost performance and robustness in real-world settings, even with notable gaps between real and sim. Check out the thread below for details ๐.
How to use simulation data for real-world robot manipulation? We present sim-and-real co-training, a simple recipe for manipulation. We demonstrate that sim data can significantly enhance real-world performance, even with notable differences between the sim and the real. (1/n)
0
0
26
GROOT N1 is out! We show how simulation and video generation models help pave the way toward building generalist robots. Check out the white paper:.
Thrilled to announce GR00T N1, our open foundation model for generalist humanoid robots!. GR00T N1 adopts a dual-system design, leverages the entire data pyramid for model training, andย supports various robot embodiments. GR00T N1 embodiesย years of fundamental research, spanning.
1
2
28
RT @texas_robotics: Exciting news! ๐ Texas Robotics faculty members @FAlambeigi and @yukez have been awarded tenure! Their groundbreakingโฆ.
0
6
0
RT @DrJimFan: We are living in a timeline where a non-US company is keeping the original mission of OpenAI alive - truly open, frontier resโฆ.
0
2K
0
RT @kevin_zakka: The ultimate test of any physics simulator is its ability to deliver real-world results. With MuJoCo Playground, weโve coโฆ.
0
185
0
Fun thought experiment. How much does it cost to collect 1B robot demonstrations?. My guess is it's well north of $100M:.* 3500 robots running 247 to collect over 1 year: Cost 70M+.* Data collection outside US Cost: 50M+. Eye watering but feasible. But is 1B demos all you need?.
5
5
42
RT @smithlaura1028: Excited to share our work on STEERing robot behavior! With structured language annotation of offline data, STEER exposeโฆ.
0
19
0
RT @JasonMa2020: Excited to finally share Generative Value Learning (GVL), my @GoogleDeepMind project on extracting universal value functiโฆ.
0
115
0
RT @xiao_ted: Can we improve VLA generalization abilities by bridging the gap between robot data and internet data?. We find that *visual aโฆ.
0
11
0
Please see the paper for more details. This was my internship project at Google DeepMind. A huge thank you to my awesome mentor @xiao_ted for supporting me and all of my lovely collaborators @SeanKirmani @TianliDing @smithlaura1028 @yukez @DannyDriess @DorsaSadigh!.
0
0
1
Hereโs the big kicker: we can adapt to new tasks and objects by just providing cheap-to-collect example images and annotating them with affordances. No additional costly robot demonstrations or teleoperation required!
1
1
3
Our hierarchical model first predicts an affordance plan and then conditions the policy on the affordance plan. We co-train the model on web datasets (largest data source), robot trajectories, and a modest number of cheap-to-collect images labeled with affordances.
1
0
1
We want to make a robotโs job easy by telling it not only what to do but how to do it. Conditioning on language, goal images, and trajectory sketches are helpful, but they present their own challenges. Visual affordance plans are expressive and easy to specify!
1
0
0
Excited to release RT-Affordance! We propose conditioning policies on visual affordance plans as an intermediate representation that allows us to learn new tasks without collecting any new robot trajectories. Website and paper: Hereโs a short ๐งต
3
29
137