Jiangran Lyu

MLLMs possess an intrinsic "S-Space" that acts as a spatial workspace, encoding object coordinates and enabling viewpoint transformations, even under non-spatial queries.

Adding just 5% of target-embodiment data during pretraining can boost transfer performance by over 13 percentage points, highlighting the critical role of embodiment exposure in VLA models.

HarnessEval-W transforms world model evaluation from mere scoring to a transparent reasoning process that mirrors human judgment.

WAM-TTT allows robot models to adapt to new tasks using only raw human videos, eliminating the need for additional demonstrations or fine-tuning.

KPGrasp achieves a remarkable 76.3% grasp success rate, outpacing existing methods by a staggering 47.4% while simplifying the grasp generation process.

Cosmos 3 sets a new benchmark for omnimodal models, outperforming existing state-of-the-art in both Text-to-Image and Image-to-Video tasks.

Generating coordinated bimanual grasps on diverse objects is now possible thanks to a dataset of nearly 10 million grasps and a model that adapts to object geometry and size.

Forget hand-crafted heuristics: this new dynamics-aware policy learns to exploit contact forces in cluttered environments, outperforming traditional methods by 25% in simulation and showing impressive sim-to-real transfer.

Training a robot foundation model on 30,000 hours of heterogeneous embodied data lets it outperform prior methods by up to 48% on complex manipulation tasks and even benefit from low-quality data.