Multi-object grasping remains a hard robotics problem because of high data-collection costs, variable spatial layouts, and the need to generalize to new objects and embodiments. The new MoDex framework, accepted at CoRL 2026, proposes a unified solution that works across diverse spatial arrangements without task-specific retraining.

MoDex simplifies the problem by treating it as a single-step decision task: rather than planning a full grasp from scratch, the policy learns residual corrections to a reference trajectory. A spatially adaptive dual-strategy then decides how to interact—performing a single-shot grasp when objects are close together, and switching to sequential interaction when they are separated.

In experiments, MoDex significantly outperformed baselines on proximal configurations, reaching 61.24% success on unseen objects, and also achieved high success rates in separated layouts. The authors report generalization across different robot embodiments and zero-shot sim-to-real transfer using a vision-based policy trained only on simulation data. Code and pretrained models are publicly available.