Four papers ease robot retraining and long-horizon skill transfer

Four recent arXiv papers converge on the same pain point: getting robot policies to handle new tasks and new bodies without starting from scratch. [1] proposes an agent-centered architecture where a robot "teaches" itself new manipulation skills by growing its own repertoire, reducing the repeated data collection and policy optimization typical of vision-language-action models. [2] instead focuses on closed-loop control code, arguing that coding agents can generate and transfer software that already includes state management and branching, cutting manual tuning effort.

Where the papers differ is in mechanism. [3] addresses the long-horizon problem with StageGuard, which learns when to switch from one skill to the next via agentic distillation, improving hierarchical planning. [4] tackles cross-embodiment transfer in latent action models, adding action-similarity supervision so that skills learned on one robot can transfer to another without dense demonstrations. These are complementary routes: one improves task sequencing, another improves action representation across morphologies.

The shared thread is a push toward modularity and reuse. All four assume that generalist robot learning will not scale if every new task or robot requires full retraining. Yet they disagree on the right abstraction—whether to grow an agent's own skills, transfer software, learn stage transitions, or refine latent action spaces. Together they suggest the field is converging on reducing human and compute overhead, even as the solutions remain fragmented.