PhD student in robotics manipulation, currently on physics-aware world models for robust manipulation @RiceCompSci. Designer and developer of XLeRobot.
Joined November 2021
- Tweets 465
- Following 240
- Followers 2,100
- Likes 989
186 Photos and videos
Pinned Tweet
Apr 30
Catching a flying ball is hard. What if with a flat plate? š
Our work at RSSā26 shows itās possible through Zero-Shot Sim2Real
With Domain-Randomized Instance Set (DRIS), we catches different kinds of balls without any real-world fine-tuning
š rice-robotpi-lab.github.io/Dā¦
10
22
171
67,745
introducing lattice
a low-cost robot arm with a modular, detachable gripper so the robot can swap its own tools
the adapter is $14 if you already have an so101 arm, and about $140 to build the full system from scratch
itās now fully open source
39
43
516
38,547
Vector Wang retweeted
May 29
Iāll be at #ICRA2026 in Vienna presenting ScheduleStream, our work on GPU-accelerated multi-arm task and motion planning: schedulestream.github.io
If youāre looking to meet to chat about NVIDIA, schedule time here: docs.google.com/forms/d/e/1Fā¦
6
12
149
13,413
May 30
Started my research internship in the NVIDIA Seattle Robotics Lab. Mostly working on agentic robotics.
In Seattle this summer and ready to hang out with new/old friends!
27
16
798
32,940
Vector Wang retweeted
š Panthera-HT is here ā the cost-efficient robotic arm that doesn't compromise performance. Only $1,799.3.5kg payload | 0.8m/s speed | ā¤140Ī¼s response Fully open-source (MIT) | LeRobot-native | ROS2 ready
#Robotics #OpenSource #EmbodiedAI #LeRobot #XLeRobot
10
16
149
17,397
May 9
This is really professional
May 9
Setting up with @IsaacSin12 for our upcoming hackathon
This is part of our @makermodsai Physical AI Hack World Tour.
Weāre introducing OpenBooth, our new foldable physical AI lab. If you want a grab one for yourself, move quick coz we are selling out fast.
This is a massive robotics hackathon here in @fdotinc. Just putting in some final few touches on our robots and OpenBooth setups for our hacker tomorrow.
2
1
18
2,296
Vector Wang retweeted
May 7
.@NVIDIARobotics is joining the SF Physical AI Hackathon as a sponsor
Every team gets Brev credits to train models on cloud GPUs.
Win the hackathon and walk away with a Jetson Orin Nano Super ! Brev credits
May 9ā10 Ā· @fdotinc Fort Mason, SF
RSVP ā luma.com/0xbcm3kj
5
9
59
10,900
Apr 30
Catching a flying ball is hard. What if with a flat plate? š
Our work at RSSā26 shows itās possible through Zero-Shot Sim2Real
With Domain-Randomized Instance Set (DRIS), we catches different kinds of balls without any real-world fine-tuning
š rice-robotpi-lab.github.io/Dā¦
10
22
171
67,745
Apr 30
We put DRIS to the test with a highly demanding task: reactive catching using a completely flat plate on a Franka Robotics FR3 @FRANKAROBOTICS robot. Unlike robotic systems that rely on cups, nets, or enclosing surfaces for passive stabilization, our setup offers zero mechanical help, making it extremely sensitive to contact timing and noise.
375
Apr 30
Accepted at Robotics: Science and Systems (RSS Conference) 2026:
"Zero-Shot Sim-to-Real Robot Learning: A Dexterous Manipulation Study on Reactive Catching."
Authors:Kejia RenĀ¹, Gaotian WangĀ¹, Andrew S. MorganĀ², Kaiyu HangĀ¹
Ā¹Rice University Ā²Robotics and AI Institute
622
Apr 30
Same as ManiDreams, this task is also fully trained within Maniskills @Stone_Tao šš
1
1
395
Apr 30
Here is how it works:Ā
Instead of training on one randomized instance at a time, DRIS propagates multiple randomized instances simultaneously under a shared action.Ā
By allowing the policy to learn from a set of possible dynamics outcomes at once, it becomes remarkably robust to variations in physical properties and sensing errors
2
499
Vector Wang retweeted
Apr 29
We're building the UI @LeRobotHF should have shipped with
Plug in your SO101 or XLeRobot, scan your ports and cameras, auto-calibrate, and you're teleoperated in under 5 minutes. No terminal.
7
12
100
8,779
Apr 28
Building up a mini Tidybot for <$2k in total.
More dynamic (realtime no speed up) and more robust
Plugged into upcoming X-bot universe with OpenClaw agentic applications.
github.com/TidyBot-Services/ā¦
16
31
333
38,223
Vector Wang retweeted
Apr 23
Watch AthenaZero juggle barehanded using on-board sensory feedback only. No motion capture. No funnels. No help adding the third ball. The robot learns to adapt to the uncertainties from contact and the appropriate hand-eye coordination.
Learn more: rai-inst.com/resources/blog/ā¦
2
20
103
13,658
Apr 20
No world model is accurate, especially the intuitive one in your head
Most video models collapse the future into one deterministic rollout, slowly
ManiDreams keeps the uncertainty, and plans over it in a modular framework
š“Dream, š¤Predict, š¦Constrain
rice-robotpi-lab.github.io/Mā¦
8
37
313
158,513
Apr 20
ManiDreams is not a standalone method, but:
An Open-Source Library for Robust Object Manipulation via Uncertainty-aware Task-specific Intuitive Physics
Authors: Gaotian WangĀ¹, Kejia RenĀ¹, Andrew S. MorganĀ², Kaiyu HangĀ¹
Ā¹Rice University Ā²Robotics and AI Institute
(Video: Fast Foundation Stereo for perception Newton for physics backend, real-time 3D DRIS physics prediction)
4
410
Apr 20
Thanks to @Stone_Tao, a majority of task evaluations, physics backends, and executors are built upon Maniskill.
3
333
Apr 20
Modular, a plug and play contract
Sampler: Heuristic, RL, MPPI, or VLA.
Physics backend: ManiSkill, Newton, or diffusion models. JEPA style models later. World models not limited.
On RTX 4090, the whole loop runs faster than real time (20Hz) for both simulation and diffusion.
3
408
Apr 20
Core concepts:
DRIS. Domain Randomized Instance Set. Perceptual, structural, and parametric uncertainty live as a set of randomized instances
TSIP. Task Specific Intuitive Physics. Forward prediction runs over DRIS
Cage. It asks "does this action keep the uncertainty bounded"
3
431
Apr 20
Drop bunnies with 1mm and 0.01rad of randomness, they land completely different. No model can tell you which one is ātheā future, because there isnāt one.
When catch a ball or pour coffee, your brain runs a blurry distribution. That blur isnāt a bug, but prediction under noise.
7
592