What Is ReconVLA?

ReconVLA is introduced in ReconVLA: Reconstructive Vision-Language-Action Model as Effective Robot Perceiver (arXiv 2508.10333) by Wenxuan Song, Ziyang Zhou, Han Zhao, Jiayi Chen, Pengxiang Ding and co-authors, submitted on 14 August 2025. The abstract starts from an empirical finding: current VLA models struggle to allocate visual attention to target regions, and their attention is often dispersed.

ReconVLA is third-party research. Firsthand is not part of it or affiliated with its authors, and makes no claim that any model was trained on Firsthand data. What the site states is custom egocentric collection of real human demonstrations, fast and on demand, to a written spec: 3D hand pose at 21 joints per hand, calibrated stereo depth and 6-DoF IMUs at 200 Hz, delivered in RLDS, LeRobot format or WebDataset with a buyer-owned commercial license. The only language data the site states is one task instruction per episode. Whether Firsthand data suits a ReconVLA-style pipeline is Not stated in the sources read.

Per the abstract, VLA models combine multimodal understanding with action execution, but the authors’ analysis shows they struggle to allocate visual attention to the regions that matter for the task. Attention spreads across the image instead of settling on the object being manipulated.

The README describes a model with a reconstructive part and an action part. The input is multi-view images and a text instruction. The action part outputs discrete action tokens. The reconstruction part outputs reconstructive tokens, which condition a denoising process that rebuilds scene tokens from noisy ones. The scene tokens are tokenized images of gaze regions.

The abstract says a diffusion transformer reconstructs the gaze region, which corresponds to the target manipulated objects. Learning to do this prompts the model to build fine-grained representations and allocate attention accurately.

The abstract says the authors curated a large-scale pretraining dataset of over 100k trajectories and 2 million data samples from open-source robotic datasets. The project page names BridgeData V2, LIBERO and CALVIN as its basis.

The README states the repository ships with no raw data. Users download those public datasets and preprocess them, including generating target images of the gaze regions with object detection and grounding methods such as GroundingDINO or YOLO.

The abstract reports experiments in simulation and the real world that show the superiority of the implicit grounding method, with precise manipulation and generalization.

These are the authors’ results on their own setups. Firsthand does not reproduce them and quotes no scores.

The official repository, OpenHelix-Team/ReconVLA, is released under the MIT License. Its README includes pretraining and fine-tuning scripts and an evaluation flow built on CALVIN.

The site’s TraceVLA entry covers visual trace prompting, which overlays point-tracked traces on the input image. Its SpatialVLA entry covers Ego3D Position Encoding and Adaptive Action Grids. Its DreamVLA entry covers forecasting of dynamic regions, depth and semantics. ReconVLA’s distinguishing part, as stated by its authors, is a reconstruction objective over gaze regions. All are third-party methods, and Firsthand ranks none of them.

This entry draws on the arXiv abstract page, the official GitHub repository and the official project page. None describes human egocentric video, stereo depth, hand pose or IMU data in ReconVLA’s training or evaluation, so those are Not stated.

When a program needs real human egocentric demonstrations that public data does not cover, specified in writing. Firsthand does not build or train models. Confirm the spec on a Spec Pilot, a small fast first batch; the median from signed spec to first validated batch is 48 hours where coverage is deep, three to four weeks where thin.

What is ReconVLA?

A published reconstructive VLA model (arXiv 2508.10333) in which a diffusion transformer reconstructs the gaze region of the image, to guide the model’s visual attention onto the target object.

What is a gaze region in ReconVLA?

Per the README, the scene tokens are tokenized images of gaze regions, and the abstract says the gaze region corresponds to the target manipulated objects.

What did the authors report?

The abstract reports that experiments in simulation and the real world show the superiority of the implicit grounding method. These are the authors’ results, and this page quotes no scores.

Does Firsthand data train ReconVLA?

No claim is made. Firsthand is not affiliated with its authors, and the site does not state that any model was trained on Firsthand data.

Can Firsthand collect egocentric data for a VLA program?

The site states custom egocentric collection to a written spec, fast and on demand, with 21-joint 3D hand pose, calibrated stereo depth, 200 Hz IMUs and one task instruction per episode, confirmed on a Spec Pilot.