What Is CronusVLA?

CronusVLA is introduced in arXiv 2506.19816 by Hao Li, Shuai Yang, Yilun Chen, Xinyi Chen, Xiaoda Yang, Yang Tian, Hanqing Wang, Tai Wang, Dahua Lin, Feng Zhao and Jiangmiao Pang (v1 24 Jun 2025, v2 30 Oct 2025). The project page lists the University of Science and Technology of China, Shanghai Artificial Intelligence Laboratory, Zhejiang University and The Chinese University of Hong Kong. Read: the arXiv abstract page, project page and repository 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: 21 keypoints per hand with a visibility flag, wrist trajectories and action segments, plus one task instruction per episode.

The abstract says VLA models built on pretrained vision-language models remain constrained by the single-frame image paradigm and do not fully use the temporal information in multi-frame histories. It adds that feeding several frames straight into a VLM backbone brings substantial computational overhead and inference latency.

Stage one is single-frame pretraining on large-scale embodied datasets with autoregressive prediction of action tokens. Stage two is multi-frame post-training, which changes the backbone’s prediction from discrete tokens to learnable features and aggregates historical information via feature chunking. The README says the 7B model is pretrained following OpenVLA and the 0.5B model following MiniVLA, using the oxe_magic_soup_plus_minus suite, then post-trained mainly on BridgeData V2 and RT-1 (fractal) data in RLDS format.

The README and project page describe 7B and 0.5B versions. Its checkpoint table lists post-training with step 7 and action chunk 16 for the 7B model, and step 4 and chunk 16 for the 0.5B model. Evaluation takes an instruction and a third-person image, with wrist-view variants for LIBERO. The README’s inference example loads a DiT-B action model, and returns 16 steps of 7-DoF actions.

The abstract reports experiments across three embodiments in simulated and real-world environments, with a 70.9% success rate on SimplerEnv, a 26.8% improvement over OpenVLA on LIBERO, and the highest robustness score on SimplerEnv-OR. The project page adds real-world tests on a Franka Research 3 robot with a third-person camera. These are the authors’ numbers. The project page’s BibTeX lists a different title from arXiv, so this page uses the arXiv title.

SimplerEnv-OR is a benchmark the authors introduce to test robustness to observational disturbances. The abstract says it features 24 types of disturbances and 120 severity levels, and the project page splits them into temporal and spatial dimensions, with examples such as occlusion, jittering and impulse noise. The project page compares several other VLAs, including CogACT, with CronusVLA on it; scores are Not stated here.

The site’s entries describe OpenVLA as a 7-billion-parameter model on a Llama 2 language model and CogACT as using a diffusion transformer action module conditioned on a cognition feature. CronusVLA builds on a single-frame VLA, following OpenVLA for 7B pretraining, and adds the multi-frame stage. The table sets them side by side; this page ranks none of them.

It means using past observations when predicting actions, not the current frame alone. The sources describe aggregating historical features through feature chunking, not passing raw frames through the whole backbone. Per-task history length is Not stated here.

The full paper’s method detail, the Hugging Face model cards and the real-world task-level results were not read. CronusVLA’s stated data is robot trajectories with third-person images and instructions, and the README says custom data can be converted to RLDS, with Franka data collected by teleoperation. The sources do not say whether first-person human video would suit it. The README’s MIT licence applies to the repository; terms for the datasets it names are Not stated here.

CronusVLA’s stated inputs are robot data, so human video is not its stated input. Firsthand captures human first-person video, not robot data: hand pose, wrist trajectories, action segments and one task instruction per episode; robot actions are Not stated. Write tasks, environments and signals into a spec and confirm on a Spec Pilot. Median time to first validated batch: 48 hours where coverage is deep, three to four weeks where thin. Formats include RLDS; matching CronusVLA’s expected fields is the buyer’s to check.

What is CronusVLA?

A vision-language-action framework (arXiv 2506.19816, AAAI26 oral) that extends single-frame VLA models to multi-frame history through single-frame pretraining and multi-frame post-training with feature chunking.

How does CronusVLA use multiple frames?

Per the abstract, post-training changes the backbone’s prediction from discrete tokens to learnable features and aggregates historical information via feature chunking, instead of feeding every frame into the backbone.

What is SimplerEnv-OR?

A benchmark introduced with CronusVLA to test robustness to observational disturbances, with 24 disturbance types and 120 severity levels, per the abstract.

Is CronusVLA open source?

The official repository states the MIT licence and links Hugging Face checkpoints from its README. The model cards were not opened here.

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

Firsthand is not affiliated and claims no use, training or compatibility. You can request a custom dataset, confirmed first on a Spec Pilot.