Fudan University

“Feng Xu, Fudan University, Corresponding author.”

“Xin Zhang, Fudan University / TeleAI intern, Co-first author.”

“TourPhysics fundamentally restructures how video world models work: instead of letting a learned video model implicitly predict physics, a deterministic simulator computes the complete physical and camera trajectory first, and only then does the video generator render an observation of that fixed trajectory.”

“Xipeng Qiu is a prominent researcher in natural language processing and foundation models, known for his work at Fudan University on large language models”

“VLA-Pro stores task-specific LoRA adapters as parameterized procedural memories during training. At inference time, VLA-Pro retrieves relevant procedural memories based on the current multi-modal context and dynamically fuses these memories for generating the current action chunk.”

“Without VLA-Pro, a state-of-the-art VLA model achieves only a 5.8% success rate on unseen real-world manipulation tasks. With VLA-Pro, that climbs to 65.0% — an 11x improvement.”

“In simulation (RoboTwin, RLBench), the framework delivers up to a 207% relative improvement.”

“the architecture can be layered onto existing investments (e.g., OpenVLA, π0, or other foundation models) rather than requiring a full model replacement.”

“Yu-Gang Jiang is a senior figure in computer vision and video understanding, and his lab's pivot into embodied AI / VLA research signals growing Chinese academic investment in Physical AI foundations.”

“His involvement suggests the efficiency angle of LoRA-based memory (small, modular adapters vs. full fine-tunes) is a deliberate architectural choice informed by efficient ML principles, not just a pragmatic shortcut.”

“Participants: Xiaosong Jia, Yu-Gang Jiang, et al. (arXiv Physical AI)”

“the depth head alone improves performance on the 'Click Bell' precision task (requiring Z-axis control) from 35% to 63%; the skill head prevents stage-collapse on long-horizon tasks; the object head raises scores on cluttered-scene tasks by +8.4% over π₀ on the LIBERO-Plus Object suite.”

“π₀'s intrinsic object-region attention mass is only 8.1%, and its argmax-hit accuracy (does the model even look at the right object?) is just 2.2%. GuidedVLA raises these to 84.0% and 84.7% respectively.”

“GuidedVLA demonstrates that VLA models fail not because they lack capability, but because their action decoders latch onto irrelevant visual noise — and fixing this with explicit, targeted supervision yields consistent 10-20% performance gains across both simulation and real hardware, without retraining the base model.”

“standard vision encoders (e.g., SigLIP) in VLA are trained with 2D supervision and lack explicit 3D awareness... The depth head improves 'Beat Hammer Block' (height alignment task) from 78% to 96%.”

“Xiaosong Jia, Fudan University (TEAI / Shanghai Key Lab of Multimodal Embodied AI). Project lead and primary technical architect of GuidedVLA. Also co-author of Interleave-VLA.”

“Zuxual Wu, Fudan University. Correspondence author. Senior PI on the Fudan side. His lab's focus on multimodal embodied AI makes GuidedVLA a flagship output for the Institute of Trustworthy Embodied AI (TEAI).”

“DeFI achieves an average task length of 4.51 on CALVIN ABC-D (the hardest generalization split, training on ABC, testing on unseen environment D), beating the prior SOTA VPP at 4.33 (+4.2%), Physical Intelligence's π0 at 3.84, and GR00T N1 at 4.01.”

“Li Zhang — Corresponding author. Oversees a lab working across video understanding and embodied AI. The institutional affiliation with the Shanghai Innovation Institute signals proximity to Chinese national AI infrastructure priorities.”