Jiageng Mao

This year, we won 2 Best Paper Awards at #CVPR2026 workshops: Humanoid Everyday (humanoideveryday.github.io/) won a Best Paper Award at the Embodied AI Workshop Ψ0 (psi-lab.ai/Psi0) won a Best Paper Award at the 3D-LLM/VLA Workshop Congratulations to the team! @zhenyuzhao123, Hongyi Jing, @SonglWei, @Boqian_Li_ , @yuewang314

(4/5) Physics-based Video Editing Once we capture the wind field, we can seamlessly insert new objects into the original video. These virtual objects interact with the real estimated wind, creating physically consistent simulations.

(3/5) We jointly model Geometry, Physics, & Interactions. By integrating a differentiable physics simulator, we use backpropagation to minimize the error between simulated and observed motion, effectively "seeing" the force field.

(2/5) Humans have "Intuitive Physics"—we watch a leaf twirl and know the wind speed. But for computer vision, this is incredibly hard. We propose an end-to-end Inverse Graphics framework that recovers complex, non-rigid force fields (like wind) purely from RGB pixels. No sensors, just vision.

"Who has seen the wind? Neither I nor you: But when the leaves hang trembling, the wind is passing through." @ScottZhiyuanGao will present our work "Seeing the Wind from a Falling Leaf" at #NeurIPS2025! We teach AI to "see" invisible force fields directly from video using differentiable physics. Project Page: chaoren2357.github.io/seeing… Paper: arxiv.org/abs/2512.00762

We release OpenReal2Sim, an open-source toolbox for real-to-sim reconstruction and robot simulation. A key difference from prior work is our focus on building an interactive digital twin from in-the-wild data — even Internet images or generated videos. Try it out: Interactive Demo: hesic73.github.io/OpenReal2S… Code: github.com/PointsCoder/OpenR…

How do we fix this? Introducing PhysAgent 🚀 – a new framework that enhances VLMs by integrating: 🔹 Vision foundation models (Depth, SAM, GroundingDINO) 🔹 A physics knowledge memory for improved reasoning 🔹 Chain-of-thought inference for self-verification PhysAgent boosts GPT-4o's PhysBench score by 18.4%, significantly improving physics-based reasoning and supporting embodied AI applications!

What did we find? 🧐 We evaluated 75 top VLMs, including GPT-4o, Gemini, and open-source models, and found: ✅ Strong commonsense reasoning but poor physical reasoning ✅ Closed-source models outperform open-source ones, but still struggle ✅ Scaling data and model size does not significantly improve physics understanding This suggests that current VLM training lacks physical priors, limiting their ability to reason about real-world interactions.