Hao Zhao
@HaoZhao_AIRSUN
Followers
608
Following
959
Media
11
Statuses
58
https://t.co/l52lTpdO16 Computer vision is good, have fun.
Tsinghua University
Joined July 2024
🚀 GaussianArt (3DV 2026) is here! A single-stage unified geometry–motion model that finally scales articulated reconstruction to 20+ parts with order-of-magnitude higher accuracy. Evaluated on MPArt-90, the largest articulated benchmark to date. Code + project page below 👇 🔗
1
23
158
If you’re excited by Tesla’s new world model, meet OmniNWM—our research take on panoramic, controllable driving world models • Ultra-long demos • precise camera control • RGB/semantics/depth/occupancy • intrinsic closed-loop rewards Arxiv: https://t.co/CKa5Bd9bqr Watch:
10
47
319
InvRGB+L (ICCV’25): inverse rendering of large, dynamic scenes from a single RGB+LiDAR sequence. We add a specular LiDAR reflectance model + RGB↔LiDAR material consistency, yielding reliable albedo/roughness, relighting, night sim, and realistic object insertion. Paper:
2
20
140
TA-VLA: Torque-aware Vision-Language-Action models We show how to inject force cues into VLAs for contact-rich manipulation. Three takeaways: ✅Where: put torque adapters in the decoder, not the encoder. ✅How: use a single-token summary of torque history. ✅Why: jointly
9
57
452
Thrilled to share that our paper "OnePoseViaGen" received three strong accepts at #CoRL2025! 🚀🎉 Check it out 👉 https://t.co/U54kgzIXGy Congrats to the amazing team! 🔥
github.com
[CORL 2025 Oral]One View, Many Worlds: Single-Image to 3D Object Meets Generative Domain Randomization for One-Shot 6D Pose Estimation. - GZWSAMA/OnePoseviaGen
Excited to share that our RGB-D version of OnePoseViaGen has been accepted to #CoRL2025 with three strong accepts! 🤖 Code coming soon at https://t.co/1D2sT9l9P2 6D pose estimation is crucial for robotics, but generalizing to generated objects remains challenging. We introduce
0
1
15
🚀 Mind-blowing! OnePoseViaGen can track 6D object poses directly from any input video – no special setup needed! From just an RGB video 🎥, it reconstructs 2D depth, 3D shape, 4D point dynamics, and 6D pose. 👉 Try it here:
huggingface.co
🚀We are thrilled to release an alpha version of OnePoseViaGen - a system for Panoptic 4D Scene Reconstruction from RGB video! 👉Try it at : https://t.co/RFKddVaVCV
#AI #3D #GenerativeAI
1
14
58
🚀 New code release! We’ve open-sourced DiST-4D — the first feed-forward world model that simultaneously handles temporal prediction and spatial novel-view synthesis for autonomous driving scenes. • Disentangled spatio-temporal diffusion • Metric-depth bridge for 4D RGB-D
0
9
21
🙌 Great to see @dreamingtulpa showcasing SyncTalk++! We’ve level-led up the original NeRF-based SyncTalk with 3D Gaussian Splatting: • Dynamic Portrait Renderer for sharper, consistent identity • FaceSync + HeadSync for spot-on lip & pose alignment • 101 fps real-time
2
16
48
Photometric stereo meets VGGT: LINO leverages geometry backbones + light register tokens to deliver universal, 4K-detailed normal maps under arbitrary lighting. 👀 Thanks for the post @zhenjun_zhao
Light of Normals: Unified Feature Representation for Universal Photometric Stereo Hong Li, Houyuan Chen, @ychngji6, @Frozen_Burning, Bohan Li, @xshocng1, Xianda Guo, Xuhui Liu, Yikai Wang, Baochang Zhang, Satoshi Ikehata, Boxin Shi, @raoanyi, @HaoZhao_AIRSUN tl;dr: learnable
0
13
46
LINO = VGGT + Learnable Light Tokens + Detail-Aware Losses 🔥 Huge thanks to @raoanyi @chen_yuan76802 — loved building this together! Project:
houyuanchen111.github.io
TWITTER BANNER DESCRIPTION META TAG
Universal Photometric Stereo (PS) aims for robust normal maps under any light. 🚨 But big hurdles remain! 1️⃣ Deep coupling: Ambiguous intensity - is it the light changing or the surface turning? 🤔 2️⃣ Detail loss: Complex surfaces (shadows, inter-reflections, fine details) stump
0
15
53
Combining VGGT with lighting registers gives rise to today’s strongest foundation model for photometric stereo. Thanks @_akhaliq for highlighting our work on LINO: predicting ultra-detailed 4K normal maps from unified features! 👀
Light of Normals Unified Feature Representation for Universal Photometric Stereo
1
10
64
Just arrived LA! We will present 4 papers @RoboticsSciSys including award candidate paper Reactive Diffusion Policy @HanXue012 , DemoGen @ZhengrongX , DoGlove @DoubleHan07 , and Morpheus @HaoZhao_AIRSUN robot face. I'll also share thoughts about OOD generalization in workshops.
1
9
75
🚗📡 Simulate Any Radar (SA-Radar) is here! We present a controllable, efficient, and realistic radar simulation system via waveform-parameterized attribute embedding. 🌐 Supports: • Cross-sensor simulation • Attribute editing • Scene augmentation • RAD cube generation 🔍
0
1
4
🤖 Just published in Nature Machine Intelligence! F-TAC Hand embeds high-res touch (0.1 mm) across 70% of a biomimetic robotic hand — enabling adaptive, human-like grasping across 600 real-world trials.
Our paper, published in Nature Machine Intelligence, presents a system with full-hand tactile sensing and sensory-motor feedback for adaptive, human-like grasping, advancing embodied agents in real-world operation. Article: https://t.co/LwN7ZzyOfx Demo: https://t.co/rhi0cztkdp
0
1
8
🤖 Meet Morpheus — a neural-driven animatronic face that doesn’t just talk, it feels. Hybrid actuation (rigid 💪 + tendon 🧵) makes it expressive and compact. Self-modeling + audio-to-blendshape = real-time emotional reactions 😮💨😠🥹 🧠💬 Watch it smile, frown, cringe... all
2
8
20
🚀 Looking forward to presenting three works at #CVPR2025 next week! Come check them out if you're interested in dynamic modeling, 3D physical reasoning, or generative driving scenes: 🔹 PartRM: Modeling Part-Level Dynamics with Large Cross-State Reconstruction Model 👉
github.com
Contribute to GasaiYU/PartRM development by creating an account on GitHub.
We are one week away from #computervision’s largest conference #CVPR2025! 🤗 What are you most excited to see?
0
1
5
🚘 Tsinghua & Bosch just dropped Impromptu VLA: the SOTA fully open-source, end-to-end Vision-Language-Action model for autonomous driving — no BEV, no planner, just raw video ➝ natural language ➝ action. Beats BridgeAD on NeuroNCAP (2.15 vs. 1.60). 👉 https://t.co/Ub2jisUnKG
0
22
101
We propose **Challenger**—a framework to generate **photorealistic adversarial driving videos**! ⚠️🚗 - 🚗 Diverse scenarios: **cut-ins, tailgating, blocking**, without human supervision - 💥 **8.6× to 26.1×** higher collision rates for SOTA AD models - 🎯 **Transferable**
1
7
22
The physical realism here is wild. PhysGen3D easily beats most closed-source stuff (Kling, Runway, Pika…) when it comes to physics-informed generation. 🚀 #CVPR2025
Can we make images interactive with realistic physics? 🚀 Thrilled to share our #CVPR2025 work: PhysGen3D! From just a single image, PhysGen3D creates an interactive, physics-informed 3D scene, enabling us to explore and simulate realistic future scenarios interactively.
1
20
69