RT @tianminshu: 🚀 Excited to introduce SimWorld: an embodied simulator for infinite photorealistic world generation 🏙️ populated with diver….

Our paper, dataset and code are all available online:.Demos & code: Dataset: arXiv: Thanks to all my awesome collaborators @ZhengtaoHan, Shivam Aarya, Jianxin Wang, Shuhang Xu, Jiawei Peng, Angtian Wang,.

Our ablation studies show that models with explicit part-level representations can outperform the vanilla models. 3D part representations seem to be more helpful in learning part-level manipulation tasks than the 2D ones. [6/n]

We evaluated several SOTA robot manipulation approaches, including end-to-end vision-language policies and hierarchical planning models for robot manipulation on PartInstruct. • We observe low success rates among all end-to-end VLA policies (<15% success). Methods struggle

We focus on generalization evaluation, including:.• Novel object positions and rotations (OS).• Novel object instances within the same category (OI).• Novel part combinations within the same task categories (TP).• Novel part-level manipulation task categories (TC).• Novel.

PartInstruct contains diverse object assets richly annotated at part-level, and a large set of expert demonstrations for training policy models:.• 513 object instances across 14 categories 📦.• 1,302 fine-grained tasks in 16 classes 📜.• 4,043 part-level task instructions 💬

Fine-grained robot manipulation requires robust reasoning about object parts and their relationships with intended tasks. To follow instructions like “pick up the bottle and show me the cap,” the robot must identify relevant parts, ground them in 3D perception, and plan

🚀New robot manipulation benchmark.How to teach robots to reason about and interact with relevant object parts for a given fine-grained manipulation task? To address this challenge, our #RSS2025paper introduces PartInstruct, the first large-scale benchmark for fine-grained