The performance of learned robot visuomotor policies depends heavily on the size and quality of their training data, yet collecting high-quality demonstrations remains costly for robots in the real world. Although large-scale robot and human datasets are increasingly available, embodiment gaps and mismatched action spaces make them difficult to leverage directly. Cross-embodiment transfer — reusing experience from other embodiments to improve learning on a target embodiment — is therefore crucial for scaling robot learning beyond per-robot data collection. This thesis argues that efficient transfer can be greatly beneficial by learning from what is shared across embodiments — the visual dynamics of how the world responds to motion — and by effectively exploiting the scarce target-embodiment data at test time.

In this thesis, a framework for efficient cross-embodiment transfer is presented for object manipulation: Latent Policy Steering (LPS). A World Model (WM) is pretrained across eight different embodiments; by leveraging optical flow as a shared representation of visual transitions, we show that such a WM can be adapted to an unseen target embodiment with only one hour of teleoperation data. To fully exploit this limited target-embodiment data, we develop a robust value function that simulates, in the WM's latent space, the test-time mistakes the target-embodiment policy is prone to make, and steers the policy back toward the training data distribution. Through evaluations on both real-world tasks and simulation benchmarks, the proposed framework consistently improves the policy across tasks and outperforms baselines that rely on embodiment-specific proprioception and actions during pretraining.

Together, the results support the two keys of this thesis. A world model grounded in embodiment-agnostic visual dynamics, rather than any particular proprioception or action space, learns transferable representations from diverse, cheaply available experience that adapt to a new embodiment from few demonstrations. And once finetuned on that scarce target data, the same model lets the policy anticipate the consequences of its actions, foresee and avoid its own mistakes through test-time search — making steering a desirable paradigm for leveraging pretrained representations.