Robots can generate continuous movements without a separate action network

The authors’ shared vision-language model generates robot trajectories directly, with action-specific projections accounting for just 0.21% of its parameters.

Top University

Chengcheng Wang · Wei He · Wenhao Li · Hongguang Li · Kai Han · Jianyuan Guo · +1 more

University of Sydney · TokenRhythm Technologies · City University of Hong Kong

Research Digest·6 November 2026·2 min read

Illustration: Robots can generate continuous movements without a separate action network

In one line

Concertina unifies manipulation and cognition in a single diffusion vision-language model via continuous latent action tokens.

Before you read this

Three ideas this paper builds on

diffusion modelA type of generative model that learns to reverse a gradual noising process to create data.178 papers →vision-language-action modelA robot policy model that maps visual observations and language instructions to actions, trained on demonstrations and often used for manipulation tasks.77 papers →audit workflowA systematic process for verifying data integrity, model behavior, and evaluation validity through checks for errors, confounds, and hidden artifacts.100 papers →

Read it as

Most robots that follow instructions use separate systems for understanding images and language and for planning movements. This split prevents the two parts from improving each other and makes training harder.

Concertina combines both in a single model. It represents each moment of a robot's motion as a continuous token inside the model and generates actions block-by-block using the same process the model uses for images and language. No extra movement-only modules are needed.

This unified model uses very few extra parameters for generating actions, only 0.21% of its total. On standard robotic task benchmarks, it succeeded in 98.15% of trials, showing that combining thinking and acting in one model works better than separating them.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

embodied vision-language model diffusion model vision-language-action model audit workflow action chunking long-horizon robot manipulation