Concertina: Embodying Manipulation and Cognition in One Diffusion Vision-Language Model
Concertina: Embodying Manipulation and Cognition in One Diffusion Vision-Language Model
This paper presents Concertina, an Embodied Vision-Language Model (eVLM) that unifies manipulation and cognition within a single diffusion backbone. The model encodes control timesteps as continuous latent action tokens and reconstructs temporally ordered action blocks using the same operator as multimodal pretraining. With only 0.21% of parameters dedicated to active action projections, Concertina achieves 98.15% average success on LIBERO, 85.9% on LIBERO-Plus (10,030 tasks), and an average chain length of 4.142 on CALVIN ABC-D. The authors validate cross-time trajectory coupling in the shared backbone layers via an Action Interface Audit. Code and models are available on GitHub and Hugging Face.
Concertina introduces an Embodied Vision-Language Model (eVLM) architecture that integrates continuous trajectory generation directly into a shared diffusion backbone, achieving 98.15% average success on the LIBERO benchmark and 85.9% on the LIBERO-Plus benchmark while utilizing only 0.21% of parameters for active action projections.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
average completed chain length on CALVIN ABC-D
4.142
View original evidence
5.1% dedicated action-expert capacity in representative continuous VLAs. Using our proposed Action Interface Audit, we confirm that cross-time trajectory coupling emerges in the middle and late layers of the shared backbone rather than at the action interfaces. On standard benchmarks, Concertina achieves 98.15% average success on regular LIBERO, 85.9% on the 10,030-task LIBERO-Plus benchmark, and 4.142 average comple
Open source S5
- code: Available Artifact link Open source S6
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
5.1% dedicated action-expert capacity in representative continuous VLAs. Using our proposed Action Interface Audit, we confirm that cross-time trajectory coupling emerges in the middle and late layers of the shared backbone rather than at the action interfaces. On standard benchmarks, Concertina achieves 98.15% average success on regular LIBERO, 85.9% on the 10,030-task LIBERO-Plus benchmark, and 4.142 average comple
Open source S5
usion VLM backbone. Each control timestep is encoded as one continuous latent action token in the backbone hidden space, and the shared Transformer reconstructs temporally ordered action blocks via the same corrupt-and-reconstruct operator used for multimodal pretraining. Active action projections account for only 0.21% of parameters in the 4.46B model with zero separate action-generator layers, compared with 16.5%-3
Open source S4
ted chain length on CALVIN ABC-D. Code: https://github.com/lose4578/Concertina Hugging Face: https://huggingface.co/collections/Traly/concertina
Open source S6
Source：arXiv Robotics — research abstracts · arxiv.org