Sim Data Scales Up, Grasping Gets Geometry-Aware, and Long-Horizon Tasks Get a New Framework

ScaleDP: Scaling Synthetic Data Generation for Diffusion Policy Training

Synthetic data generation for robot policy training has been a bottleneck, and this paper tackles it head-on with a pipeline that scales dramatically without manual annotation overhead. If this holds up empirically, it meaningfully lowers the cost of training manipulation policies from scratch.

GeomGrasp: Geometry-Aware 6-DoF Grasp Pose Estimation for Novel Objects

Most grasp planners struggle with unseen objects, and GeomGrasp uses explicit geometric reasoning to generalize across novel shapes without retraining. Better 6-DoF grasp generalization is a core unlock for real-world deployment in unstructured environments.

LHPF: A Long-Horizon Planning Framework for Robot Manipulation via Task Graph Decomposition

Long-horizon manipulation is where most learned policies fall apart, and this task-graph decomposition approach gives robots a structured way to chain sub-tasks without losing context. The framework targets the exact failure mode that keeps robot autonomy confined to short, scripted sequences.

UniDex-V2: Unified Dexterous Manipulation with Vision-Language-Action Models Across Hand Morphologies

Teaching dexterous manipulation policies that transfer across different hand hardware is a practical problem every robot builder hits eventually, and this work attacks it directly with a unified VLA approach. Cross-morphology generalization could reduce the per-hardware retraining tax significantly.

Active Tactile Sensing for Contact-Rich Assembly Tasks Using Model Predictive Control

Tactile feedback during contact-rich assembly is one of the clearest gaps between industrial robots and human assemblers, and this paper pairs active tactile sensing with MPC to close that gap in practice. It is a tactical building block for anyone working on precision assembly or insertion tasks.

RoboEval: A Benchmark for Evaluating Foundation Models in Robotic Manipulation

Benchmarks shape what the field optimizes for, and RoboEval is a direct attempt to give foundation model evaluation in manipulation a standardized, reproducible grounding. Worth tracking if you are comparing VLAs or deciding which backbone to build on.