Embodiment-Free Pre-Training

Embodiment-free pre-training is a stage of robot policy training in which a model learns general action-generation capabilities from large quantities of manipulation data that were recorded without binding to a particular robot's body, joint layout, or control interface.

The mechanism depends on data collected via embodiment-agnostic capture devices such as hand-held or stationary gimbals that record gripper motion, scene state, and optionally video, but do not assume a fixed robot morphology. Because no specific kinematic chain is implied, the same trajectories can be reused across many downstream robot types. Models are usually pre-trained with autoregressive action generation or flow-matching objectives conditioned on language descriptions of scene state transitions, which are produced automatically by a vision-language model annotating short clips. The hope is that the resulting representations transfer cleanly when the model is later aligned to a concrete embodiment during post-training.

The principal benefit is data scalability: embodiment-bound datasets are bottlenecked by the cost and time of operating real robots, while embodiment-free capture can in principle scale to thousands of hours and thousands of scenarios. The principal cost is that embodiment-free trajectories carry less embodiment-specific signal — dynamics, contact response, joint limits — so the model must later invest in alignment to recover them. Whether the gains from scale at pre-training time outweigh the cost of closing the embodiment gap at post-training is the central empirical question.

It is still unclear how general embodiment-free pre-training is across robot morphologies with very different dynamics, such as dexterous hands versus wheeled mobile manipulators. The long-tail of rare but safety-critical manipulations may be underrepresented when data is collected without embodiment constraints. Open benchmarks that isolate the contribution of embodiment-free data from confounders such as language supervision quality remain scarce.