Quantization-Free Autoregressive Action Transformer

Psenka et al., 2023), which will be discussed in the next two paragraphs. Current transformer-based imitation learning approaches introduce discrete action representations Existing autoregressive policies, on the one hand, sidestep and train an autoregressive transformer decoder the challenge of learning in a continuous domain by discretizing on the resulting latent code. However, the initial the actions (Lee et al., 2024; Shafiullah et al., quantization breaks the continuous structure of the 2022). This discretization can introduce several drawbacks: action space thereby limiting the capabilities of It discards the inherent structure of the continuous space, the generative model. We propose a quantizationfree increases complexity by adding a separate quantization step, method instead that leverages Generative and may limit expressiveness or accuracy when fine-grained Infinite-Vocabulary Transformers (GIVT) as a direct, control is required.