What Is ObjectVLA?

ObjectVLA is introduced in ObjectVLA: End-to-End Open-World Object Manipulation Without Demonstration (arXiv 2502.19250; submitted 26 Feb 2025, last revised 28 Feb 2025 as v2, cs.RO and cs.CV) by Minjie Zhu, Yichen Zhu and colleagues; the project page lists Midea Group, East China Normal University and Shanghai University. Read: the arXiv abstract page, the arXiv HTML text of the paper, the official project page and the GitHub organisation objectvla, whose only repository, objectvla/objectvla.github.io, appears to hold the project page source.

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 or image-text pairs with bounding boxes, fast, on demand and to a written spec: 21 keypoints per hand with a visibility flag, wrist trajectories and ordered action segments, plus one task instruction per episode.

The abstract says imitation learning typically relies on large amounts of human demonstration data, which limits scalability. Its framing problem is object generalization: a robot trained to perform a task with one object struggles to transfer the skill to a semantically similar but visually different object. The authors propose ObjectVLA, a VLA approach that uses vision-language pair data as a lightweight, scalable way to inject knowledge about a target object and establish an implicit link between the object and the desired action. That is what the phrases ObjectVLA open-world object manipulation and VLA without demonstrations refer to: no demonstrations for each new object, not no demonstrations at all.

Per the paper text, the authors co-finetune teleoperated robot interaction data with a curated set of image-text pairs. Localization metadata, such as bounding boxes, bridges the two: the image-text data asks for an object’s bounding box, and the robot data is enriched with localization reasoning that is generated before each action. The paper reports a 10:1 robot-to-image-text data ratio, and that its base model is DiVLA with a Qwen2-VL-2B backbone. A second route is a smartphone: a few captured images plus brief fine-tuning of the pretrained model, per the abstract.

The abstract says ObjectVLA generalizes across 100 novel objects with a 64% success rate in selecting objects not seen during training. The project page says an ablation without bounding boxes reached only 19% on out-of-distribution objects, and that the method transfers skills to more tasks, including pick and place, push and rotate. These are the authors’ claims, not Firsthand results, and they come from the authors’ own robot setup. This page repeats no other scores.

The site’s GraspVLA entry describes a grasping model pretrained on the SynGrasp-1B synthetic dataset and co-trained with Internet grounding data. The site’s DexGraspVLA entry describes a hierarchical framework in which a vision-language model plans and a diffusion controller drives a multi-fingered hand. The site’s co-training entry defines training one policy on more than one data source at the same time. ObjectVLA uses co-training as its mechanism, aimed at objects absent from the robot data. The sources read do not compare ObjectVLA with GraspVLA or DexGraspVLA, and the table ranks none of them.

The paper’s limitations section says the image-text data was collected by the authors with a robot-mounted camera or a smartphone, that internet-sourced image-text data was not explored, and that the method struggles with novel backgrounds and lighting. Code, checkpoint availability and licence are Not stated: the GitHub organisation shows one repository, listed as CSS, HTML and JavaScript. This page does not repeat the paper’s tables or appendix details.

ObjectVLA’s sources describe two data types: teleoperated robot data on one robot platform, and authors-collected image-text pairs with bounding boxes. Firsthand’s site states human video with depth derived from a calibrated stereo rig, 21 keypoints per hand with a visibility flag, wrist trajectories, and ordered, timestamped action segments within each episode. The site’s only language data is one task instruction per episode. Robot data, image-text pairs and bounding-box labels are not stated deliverables.

ObjectVLA’s sources describe a robot policy trained on robot data plus image-text data. 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 object-generalization 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 ObjectVLA?

A published vision-language-action method (arXiv 2502.19250) that uses vision-language pair data so a robot can handle novel objects without explicit human demonstrations for each new target object.

Does ObjectVLA really work without demonstrations?

Not entirely. The abstract says no explicit demonstrations are needed for each new target object. The paper text describes co-finetuning with teleoperated robot data for the objects the robot was trained on.

What is ObjectVLA vision-language co-training?

Per the paper text, finetuning on robot interaction data and image-text pairs together, with bounding boxes linking the two, at a stated 10:1 robot-to-image-text ratio.

How is ObjectVLA different from GraspVLA and DexGraspVLA?

They are different papers, and the sources read do not compare them. ObjectVLA targets object generalization via co-finetuning; GraspVLA uses synthetic pretraining; DexGraspVLA is a hierarchical dexterous grasping framework.

Is ObjectVLA code available?

The GitHub organisation objectvla shows one repository, objectvla.github.io, listed as CSS, HTML and JavaScript. Code, checkpoint availability and licence are Not stated.

Does Firsthand use ObjectVLA, 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; image-text pairs and bounding-box labels are not stated deliverables. You can request a custom dataset, confirmed first on a Spec Pilot.