OpenHelix: A Short Survey, Empirical Analysis, and Open-Source Dual-System VLA Model for Robotic Manipulation

TL;DR

OpenHelix surveys dual-system Vision-Language-Action architectures and provides an open-source, low-cost VLA model for robotic manipulation. By standardizing MLLM (LLaVA1.0) and policy (3DDA) setups and evaluating training/integration strategies under CALVIN environments, it isolates the impact of latent representations, prompt-tuning, and projector alignment. A simple two-stage, prompt-tuned dual-system with auxiliary reasoning tasks demonstrates strong performance and practical feasibility, highlighting the importance of coupling strategies and asynchronous inference considerations. The work emphasizes reproducibility and community-driven development toward real-world robotic deployment, while acknowledging current limitations and ongoing open-source efforts.

Abstract

Dual-system VLA (Vision-Language-Action) architectures have become a hot topic in embodied intelligence research, but there is a lack of sufficient open-source work for further performance analysis and optimization. To address this problem, this paper will summarize and compare the structural designs of existing dual-system architectures, and conduct systematic empirical evaluations on the core design elements of existing dual-system architectures. Ultimately, it will provide a low-cost open-source model for further exploration. Of course, this project will continue to update with more experimental conclusions and open-source models with improved performance for everyone to choose from. Project page: https://openhelix-robot.github.io/.

Figures (7)

• Figure 1: Key Design of Dual-System VLAs. It mainly includes: MMLM Selection, Policy Selection, Latent Feature Representation Selection, MLLM Training Strategy, Policy Training Strategy, Dual-System Integration Strategy, and Dual-System Asynchronous Strategy.
• Figure 2: Three Different Evaluation Environments.
• Figure 3: Three Different MLLM Training Strategy.
• Figure 4: Evaluations on hierarchical inference. We evaluate the performance of the dual system on the CALVIN benchmark, with inference steps set to 1 and 60, respectively."Steps" refers to the inference steps of action policy during a single MLLM inference step. The longest environmental steps of the action policy ke20243d are 60, which means MLLM only inference once and represents the most typical asynchronous scenarios.
• Figure 5: Evaluation on the shortcoming of existing dual systems. From top to bottom, the first row displays the input to the MLLM. The second row visualizes a special scenario where, at environment step 3, the blue block is manually shifted to the left. In the third row, we present the top 10 words that are semantically closest to the latent embedding. The bottom row illustrates the probability distribution of spatial words associated with the latent embedding.