Introducing RTFM (Real-Time Frame Model): a highly efficient World Model that generates video frames in real time as you interact with it, powered by a single H100 GPU. RTFM renders persistent and 3D consistent worlds, both real and imaginary. Try our demo of RTFM today!

Joint work with amazing team of Yuliang Guo (Bosch Research North America), Zhihao Zhang (MSU), Xinyu Huang (Bosch Research North America), Liu Ren (Bosch Research North America) and Xiaoming Liu (MSU) (4/N)

Today at #NeurIPS2025, we present Titans, a new architecture that combines the speed of RNNs with the performance of Transformers. It uses deep neural memory to learn in real-time, effectively scaling to contexts larger than 2 million tokens. More at: goo.gle/3Kd5ojF

I think that JiT (arxiv.org/abs/2511.13720) might have been my favorite paper of 2025. From the discussions with my friends, it got quite some controversy with many people dismissing it as some trivial reinvention of x-prediction, so I would like to put my perspective on it here

Studying generalist reward models is hard: robot datasets focus on successful demos, not failures. We introduce: - a large-scale reward modeling benchmark - a data augmentation scheme - a generalist reward model that outperforms frontier VLMs Paper: arxiv.org/abs/2601.00675

👀 Found an interesting signal hidden in the evolution of #GR00t VLA pre-training data: Facts • Human videos removed after N1.5 • World-model generated data (DreamGen) removed in N1.6 Takeaway? So far, neither matches real, targeted robot data at scale. 👉 Data quality

Really amazing to see policy, world model, and value function unified in 1 model, with certain SoTA performance being shown in practice. Can not wait to check whether the three are truly aligned at deployment, and how such alignment affects the final performance in real world

It’s truly an honor to co-organize such an exciting workshop at the upcoming CVPR. Huge thanks to our co-organizers for making this happen, and sincere appreciation to our all-star speakers for accepting our invitations.

It’s great to see increasing appreciation for 360° video generators. In fact, they even offer additional benefits under controlled trajectories: complete scenes can be generated with significantly simplified trajectories, reducing the need for long paths and mitigating the

Check out new work from Entong Su for RL finetuning pre-training flow policies with residual flow steering. The motivation is simple - steering input diffusion noise can struggle to handle higher dexterity problems like multi-fingered hands, because the base policy may not cover

In my recent blog post, I argue that "vision" is only well-defined as part of perception-action loops, and that the conventional view of computer vision - mapping imagery to intermediate representations (3D, flow, segmentation...) is about to go away. vincentsitzmann.com/blog/bitter_le…

As a 3D vision researcher, it is indeed so painful to realize : Robot intelligence may not benefit from explicit 3D representations. Besides Vincent Sitzmann's deep insights, in practice, there could be two additional bottlenecks in adopting explicit 3D in perception–action

Code dropped. See below in comment!

🚀 Excited to share our latest ICLR 2026 work 3DGEER (3D Gaussian Exact and Efficient Rendering) — now open sourced! 🔗 Code github.com/boschresearch/… 🔗 gsplat integration github.com/boschresearch/…