Forward Dynamics Models

Zipei Ma, Junzhe Jiang, Jiahui Zhang, Juntong Zhang, Chun Gu, Bozhou Zhang, Jiankang Deng, Xiatian Zhu, Li Zhang

arXiv Project Code

• TL;DR: We introduce FDM, which uses action-conditioned future prediction to supervise robotic policies during training and skips future generation at inference, achieving 99.4% average success on LIBERO and absolute gains of up to 13% on real-world tasks.

TacPAC: Tactile Prediction and Real-Time Action Correction in World-Action Models for Contact-Rich Manipulation

Zipei Ma, Xiaofei Wei, Junzhe Jiang, Shunlin Lu, Li Zhang

arXiv Code

• TL;DR: We introduce TacPAC, which uses cached tactile predictions to interpret incoming touch and correct unexecuted actions in real time, improving average success from 22% to 64% across five contact-rich real-robot tasks.

$\mathcal{N}_0$-TWAM: Scaling Tactile-Native World Action Model for Contact-Rich Manipulation

NeoteAI Team & Fudan TEAI Team

arXiv Project Code

• TL;DR: We introduce $\mathcal{N}_0$-TWAM, a tactile-native world-action model that jointly predicts future vision, touch, and actions, enabling stronger contact-rich manipulation through large-scale visuo-tactile pretraining.

$\mathcal{N}_0$-Foundation: Towards the Age of Tactile Intelligence

NeoteAI Team & Fudan TEAI Team

arXiv Project Dataset

• TL;DR: We present $\mathcal{N}_0$-Foundation, a tactile-centric foundation that unifies scalable sensing hardware, large-scale multimodal data, transferable tactile representations, and standardized real-world and simulated evaluation.

DriveWeaver: Point-Conditioned Video Inpainting for Controllable Vehicle Insertion in Autonomous Driving Simulation

Junzhe Jiang, Zipei Ma, Zijie Pan, Li Zhang†

arXiv Code

• TL;DR: We propose DriveWeaver, a point-cloud-conditioned video inpainting framework that inserts controllable, temporally consistent vehicles and extracts their 3D Gaussian representations for real-time autonomous driving simulation.

Metis: A Generalizable and Efficient World-Action Model for Autonomous Driving and Urban Navigation

Jingyu Li*, Zhe Liu*, Dongnan Hu, Junjie Wu, Zipei Ma, Wenxiao Wu, Chao Han, Zhihui Hao, Zhikang Liu, Kun Zhan, Jiankang Deng, Xiatian Zhu, Li Zhang†

arXiv Code

• TL;DR: We introduce Metis, an efficient world-action model that decouples video generation and action prediction through specialized transformer experts and asymmetric attention, retaining world-model supervision during training while bypassing video generation at inference.

From Imagined Futures to Executable Actions: Mixture of Latent Actions for Robot Manipulation

Yajie Li*, Bozhou Zhang*, Chun Gu, Zipei Ma, Jiahui Zhang, Jiankang Deng, Xiatian Zhu, Li Zhang

arXiv Project Code

• TL;DR: We introduce MoLA, which transforms imagined future videos into executable representations through a mixture of pretrained inverse dynamics models, achieving consistent gains in simulation and real-world robot manipulation.

Reinforcing Action Policies by Prophesying

Jiahui Zhang*, Ze Huang*, Chun Gu, Zipei Ma, Li Zhang†

arXiv Project Code

• TL;DR: We introduce ProphRL, which learns a reusable action-to-video world model and a flow-aware RL post-training strategy to improve VLA policies efficiently, delivering strong gains in both simulation and real-robot tasks.

BézierGS: Dynamic Urban Scene Reconstruction with Bézier Curve Gaussian Splatting

Zipei Ma*, Junzhe Jiang*, Yurui Chen, Li Zhang†

arXiv Code

• TL;DR: We propose a method that models object motion using Bézier curves, achieving accurate reconstruction and effective foreground-background separation.