What Is GraspVLA?

GraspVLA is introduced in GraspVLA: a Grasping Foundation Model Pre-trained on Billion-scale Synthetic Action Data (arXiv 2505.03233) by Shengliang Deng, Mi Yan and colleagues at Galbot, Peking University, The University of Hong Kong and the Beijing Academy of Artificial Intelligence. The project page and the GitHub README list CoRL 2025. The abstract asks whether vision-language-action (VLA) models can be trained entirely with large-scale synthetic action data. The authors curate SynGrasp-1B, a billion-frame robotic grasping dataset generated in simulation with photorealistic rendering and extensive domain randomization, and present GraspVLA as a foundation model for grasping tasks. It combines autoregressive perception and flow-matching action generation in one Chain-of-Thought process, so it can be trained jointly on synthetic action data and Internet semantics data. The abstract says this helps mitigate sim-to-real gaps and supports open-vocabulary grasping. This page draws only on the paper, the project page, the GitHub repository and the Hugging Face model and dataset pages. It describes simulated robot data, not recordings of people.

Per the paper, the pipeline starts from the LVIS subset of Objaverse, filters out unsuitable categories such as weapons, and keeps 240 categories and 10,680 object instances. Objects are randomly scaled and dropped in varied poses onto a table. A grasp synthesis algorithm generates stable antipodal grasps (the project page names BoDex), CuRobo plans collision-free trajectories to reach and lift the object, and every candidate trajectory is validated in the MuJoCo physics simulator. Isaac Sim then renders RGB images from two viewpoints with randomized lighting, backgrounds, materials and camera settings. The paper says generating the dataset took 160 NVIDIA 4090 GPUs for 10 days. The appendix gives 10 million trajectories of about 100 frames each, stored at 10 Hz, a generation cost it puts at around US$5,000, and annotations for camera calibration, bounding boxes and the 3D poses of the target object and gripper. Data generation and evaluation use a Franka Panda arm with front and side views.

The model pairs a vision-language backbone with an action expert. Per the paper, the language model is a trainable InternLM2 1.8B, the vision encoder fuses features from frozen DINO-v2 and SigLIP, and a trainable projector maps vision features into language space. A conditional flow-matching action expert generates end-effector action chunks.

The connecting mechanism is Progressive Action Generation (PAG). For both Internet grounding data and synthetic data, the model first predicts a 2D bounding box of the target object. For synthetic data it then predicts a grasp pose in the robot base frame and finally an action chunk. Internet data (the paper uses GRIT) supervises only the bounding-box step, while synthetic data supervises the whole chain. The authors say this lets grasping transfer to objects that appear only in the Internet data.

The appendix lists a batch size of 384, a learning rate of 1.6e-4, and consistently high success after 120k steps. The GitHub README reports about 200 ms inference latency using roughly 9 GB of GPU memory on a single NVIDIA RTX L40s GPU.

The real-world setup uses a Franka Panda arm with two Intel RealSense cameras in a 40 cm x 50 cm x 20 cm workspace. Objects come from synthetic categories (present in SynGrasp-1B) and web categories (present only in the Internet grounding data). Each group has five test sets (basic, lighting, background, distractors and height) of 15 objects with two trials each, and a trial succeeds if the target is grasped within three attempts. The paper reports around 90% on all test sets, with comparable results on synthetic and web categories, and describes direct sim-to-real transfer.

On LIBERO, evaluated zero-shot on the Long, Goal and Object suites with tasks reformulated as pick up an object, the paper reports 82.0, 91.2 and 94.1, and says this surpasses fine-tuned pi0 and OpenVLA baselines. Against AnyGrasp, AnyGrasp is higher on common objects (100.0 versus 93.3) and faster (37 Hz versus 5 Hz), while GraspVLA is far ahead on transparent objects (86.6 versus 10.0).

For few-shot post-training, the authors collect 100 demonstrations for each of two tasks and 10 per bottle for a third, covering rare industrial components, grasping a mug without touching its interior, and sequential grasping in a dense scene. Task 1 reaches a 90% overall success rate with bounding-box annotations only.

Section 7 of the paper says data generation and evaluation use only the Franka Panda arm with front and side views, and leaves other robots and camera configurations to future work. The appendix reports adapting to a new camera or arm with 5k additional synthetic trajectories. The authors say GraspVLA struggles with ambiguous instructions such as pick up food, and that grasp labels are synthesized using force closure, which does not account for deformability. The model focuses on grasping, with pick-and-place and pushing left as future work, and PAG adds latency that the authors say may not suit fast-moving objects.

They are different kinds of data. SynGrasp-1B is simulated robot data: planned end-effector trajectories, grasp poses and bounding boxes produced by software, with rendered images from two viewpoints. The paper frames real-world robot data collection through teleoperation as labor-intensive and costly, and synthetic data as the cost-effective alternative. Firsthand captures real human egocentric demonstrations from a head-mounted camera and two wrist cameras. Nothing on this site describes simulation data, teleoperation data, robot action or joint data, grasp labels or grasp poses, and none of that is claimed here.

The site’s existing position on synthetic data is that it fills the middle of a robot data mixture and carries a sim-to-real gap in physics, lighting and contact, and that real egocentric data supplies the real-world layer a simulator can only approximate. That is a general statement about data mixtures. Firsthand makes no claim that its data trained, was used with or is compatible with GraspVLA, SynGrasp-1B or any model named here, and the two are not presented as substitutes.

Firsthand collects real egocentric demonstrations on demand, to a written spec, rather than selling a fixed catalog. The rig is a head-mounted global-shutter camera plus two wrist cameras, with active-stereo depth and 3x IMU on one hardware-triggered PTP clock. Episodes carry 3D hand pose at 21 joints per hand with a per-joint visibility flag, wrist trajectories, contact events, frame-accurate action segments and one task instruction per episode, which is the only language field the site states.

You write the spec, confirm it on a Spec Pilot, and pay per validated hour. The site states a 48-hour median turnaround where coverage is deep and three to four weeks where it is thin. Delivery is RLDS, LeRobot, WebDataset or HDF5, with a buyer-owned commercial license. Request a custom dataset below.

What is GraspVLA?

A vision-language-action model for robotic grasping, pretrained on the simulated SynGrasp-1B dataset and co-trained with Internet grounding data. It is introduced in arXiv 2505.03233 and listed at CoRL 2025.

What is SynGrasp-1B?

A billion-frame robotic grasping dataset generated in simulation with photorealistic rendering and domain randomization. The appendix gives 10 million trajectories of about 100 frames, 240 object categories and 10,680 instances.

Does GraspVLA use any real data?

Its pretraining uses synthetic action data plus Internet grounding data, per the paper. The real-world evaluations and the few-shot post-training experiments use a physical Franka Panda arm and demonstrations collected for those tasks.

How is GraspVLA different from OpenVLA?

The paper lists OpenVLA among VLA models that build on pre-trained vision-language models, and compares against it as a baseline fine-tuned on SynGrasp-1B. GraspVLA’s own pretraining is exclusively synthetic action data co-trained with Internet grounding data.

Can GraspVLA or SynGrasp-1B be used commercially?

Not stated in the sources read. The Hugging Face model card README is empty and the GitHub README License heading shows no terms. Check the repositories directly; this is not legal advice.

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

Firsthand is not affiliated with GraspVLA’s authors and makes no claim of use, training or compatibility. Firsthand data is real human egocentric capture, not synthetic. You can request a custom egocentric dataset: write a spec, confirm it on a Spec Pilot, pay per validated hour.