LAST: New method aligns vision-language with robot actions via Lie algebra

A team of researchers from multiple institutions has published LAST (Lie-algebraic Action Space Tokenizer), addressing a fundamental challenge in Vision-Language-Action (VLA) learning: the mathematical incompatibility between semantic embeddings (linear, isotropic) and robotic action spaces (non-Euclidean, anisotropic). They formalize this as a Gromov-Wasserstein alignment problem, where the goal is to make the relational geometry of action representations compatible with the semantic geometry of VL embeddings.

The solution involves a two-stage transformation. First, global topological linearization maps the action manifold into a Lie algebra, converting trajectories into fixed-length additive representations. Second, local metric discretization hierarchically breaks these representations into schemas and whitened residuals, creating approximately isotropic local charts that align statistically with the semantic metric. The authors demonstrate that this structural alignment at both global and local levels yields VLA models with superior convergence and generalizability, potentially enabling more robust robot control from natural language and visual input.