Pretrained VLA
A vision-language-action model that has been pretrained on large-scale robot demonstration data or cross-embodiment data, internalizing substantial action priors. When integrating world models, it favors decoupled designs to avoid representational conflicts.
Timeline (last 90 days)
For policies built on standard VLMs, introducing world modeling brings stable gains; for pretrained VLAs, the decoupled design is clearly more compatible.
The parameters of pretrained VLAs are already highly specialized for action generation; if the same backbone is made to simultaneously handle future prediction and action generation, the gradient directions may conflict with each other, causing the representation to be pulled apart.
All Facts (2)
For policies built on standard VLMs, introducing world modeling brings stable gains; for pretrained VLAs, the decoupled design is clearly more compatible.
50%UnverifiedThe parameters of pretrained VLAs are already highly specialized for action generation; if the same backbone is made to simultaneously handle future prediction and action generation, the gradient directions may conflict with each other, causing the representation to be pulled apart.
50%