Open Models Almanac
Methodstraining objectivelanguage modelling objective

single mention · training objective

Autoregressive discrete-action VLA pre-training

This method adapts a vision-language backbone into a discrete autoregressive robot policy.1 It uses a single next-token prediction objective across text, vision, state, and action targets, without introducing a separate continuous action head.2

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. MolmoAct2-Pretrain adapts the Molmo2-ER vision-language backbone into a discrete autoregressive robot policy while keeping the Molmo2 token interface intact — MolmoAct2 Action Reasoning Models for Real-World Deployment
  2. Together, these choices preserve a single next-token prediction objective across text, vision-language, state, and action targets, without introducing a separate continuous action head at this stage — MolmoAct2 Action Reasoning Models for Real-World Deployment

Also called Pre-training.

Filed under training objective › language modelling objective