Learning Action Priors for Cross-embodiment Robot Manipulation

Researchers have proposed new approaches to improve robot manipulation and lifelong learning, including pre-training action modules with motion priors and introducing dynamic Mixture of Experts (MoE) structures.

A recent study published on arXiv^[1] introduced a two-stage training framework for robot manipulation, pre-training the action module with motion priors before cross-modal Vision-Language-Action (VLA) alignment. This approach achieved faster convergence, higher success rates, and stronger performance on data-scarce real-world tasks. Another paper, also on arXiv^[2], proposed LiMoDE, a novel two-stage learning scheme for lifelong robot manipulation, which learns prior knowledge in the multi-task pre-training stage and demonstrated superior performance and strong lifelong adaptation in both simulated and real-world tasks. Meanwhile, a new benchmark called WatchAct was introduced^[3] to evaluate robot manipulation tasks grounded in observed human behavior, comprising 3,000 long-horizon instances across 14 tasks in four capability domains. However, current systems remain far from solving WatchAct, with the best pipeline reaching only 16.3% Success Rate in simulation^[3].