What Is OpenHelix?

OpenHelix is introduced in OpenHelix: A Short Survey, Empirical Analysis, and Open-Source Dual-System VLA Model for Robotic Manipulation (arXiv 2505.03912; submitted 6 May 2025 as v1, cs.RO and cs.CV) by Can Cui, Pengxiang Ding and 11 co-authors; the project page lists Westlake University, Zhejiang University, Xi’an Jiaotong University and HKUST(GZ). Read: the arXiv abstract page, the official project page and the GitHub README. The full paper text was not read.

Third-party research; Firsthand is not affiliated and claims no use, training or compatibility with it. Firsthand captures first-person human video, not robot data, fast, on demand and to a written spec: depth derived from a calibrated stereo rig, 21 keypoints per hand with a visibility flag, wrist trajectories and ordered action segments, plus one task instruction per episode.

The abstract says dual-system VLA architectures have become a hot topic in embodied intelligence research, but there is not enough open-source work for performance analysis and optimization. The paper therefore does three things: it summarizes and compares the structural designs of existing dual-system architectures, runs systematic empirical evaluations of their core design elements, and provides a low-cost open-source model. The GitHub README calls the repository a re-implementation of Helix. That is what OpenHelix dual-system VLA and open-source dual-system vision-language-action model refer to: a high-level multimodal language model paired with a low-level policy.

Per the project page, OpenHelix bridges the high-level multimodal language model and the low-level policy with a learned token and a linear projection. Prompt tuning trains it efficiently without altering the language model’s parameters. An auxiliary task makes the language model predict actions directly from its latent embeddings, so that it performs multimodal reasoning. The policy adopts a diffusion-based learning mechanism conditioned on visual, proprioceptive and goal features. This is how the authors describe OpenHelix for robotic manipulation.

The project page reports that the current multimodal language model is insensitive to environmental changes, which the authors call unexpected. Their analysis of action token embeddings says the embeddings mainly reflect instruction semantics, with little adaptation to visual input: the word right consistently scores higher similarity than left, regardless of actual motion. They conclude that current dual-system designs are flawed, because the model’s visual guidance is not meaningfully transferred to the action policy, and they raise the question of whether a single language model could match the performance. These are the authors’ claims, not Firsthand findings.

The abstract and project page, as read, give no scores. The GitHub README, which is later and more detailed than the arXiv v1, describes CALVIN ABC-D evaluation and lists the best OpenHelix row at 4.08 average length (prompt-tuned language model, auxiliary tasks, pretrained policy, asynchronous 10-step inference, episode length 360), against 3.27 for the policy alone. The README says OpenHelix reaches state-of-the-art performance among dual-system VLA models. Treat these as the authors’ own claims. The repository is MIT-licensed, written in Python and Shell, and the README says checkpoints were released on Hugging Face; this page links no checkpoint, because none was verified.

The site’s Hi Robot entry describes a hierarchical system in which a high-level vision-language model outputs a simple language command and a low-level VLA policy turns it into actions. The site’s HybridVLA entry describes a single large language model that combines diffusion and autoregressive action prediction. The site’s vision-language-action model entry defines the general model class. OpenHelix is a survey, an empirical study and an open-source model of the dual-system pattern. The sources read do not compare OpenHelix with Hi Robot or HybridVLA, and the table ranks none of them.

The sources read do not state the base model sizes, the training data, the real-robot results, the survey’s list of compared architectures or the appendix details, and this page repeats none of them. The README TODO list shows real-robot and humanoid deployment as unchecked items, so none is claimed here. Details beyond the abstract, project page and README are Not stated in the sources read.

OpenHelix’s README prepares CALVIN play demonstrations and language annotations for training; it states no human first-person video. Firsthand records first-person human video to a written spec: depth from a calibrated stereo rig, 21 keypoints per hand with a visibility flag, wrist trajectories, ordered action segments and one task instruction per episode. Whether such data suits a given dual-system VLA is the buyer’s to check; the sources do not say. Confirm tasks and signals on a Spec Pilot, and use fast, on-demand custom collection when a task needs its own spec. Median time to first validated batch: 48 hours where coverage is deep, three to four weeks where thin. Formats include RLDS, LeRobot format and WebDataset.

What is OpenHelix?

A published open-source dual-system vision-language-action model and short survey (arXiv 2505.03912) that compares existing dual-system designs, evaluates their core elements empirically and releases a low-cost model.

Is OpenHelix open source?

The GitHub repository OpenHelix-Team/OpenHelix is listed under the MIT License. The README says checkpoints were released on Hugging Face; this page links none, because none was verified.

How does OpenHelix connect its two systems?

Per the project page, a learned token and linear projection bridge the multimodal language model and a diffusion-based policy, with prompt tuning and an auxiliary action-prediction task.

How is OpenHelix different from Hi Robot and HybridVLA?

They are different works, and the sources read do not compare them. OpenHelix is a survey, study and open model of the dual-system pattern; Hi Robot is a hierarchical prompt-handling system; HybridVLA combines diffusion and autoregression in one model.

What did the OpenHelix study find about the language model?

The authors report that it is insensitive to environmental changes and that its action token embeddings mainly reflect instruction semantics. These are their claims, not Firsthand findings.

Does Firsthand use OpenHelix, and can I request a custom dataset?

Firsthand is not affiliated and claims no use, training or compatibility. The site states one task instruction per episode. You can request a custom dataset, confirmed first on a Spec Pilot.