What Is SmolVLA?

The authors argue that existing vision-language-action (VLA) models are typically massive, often with billions of parameters, which leads to high training costs and limited real-world deployability. They also say those models rely on academic and industrial datasets and overlook community-collected data from affordable robotic platforms. SmolVLA is their answer: a small, efficient VLA designed to be trained on a single GPU and deployed on consumer-grade GPUs or even CPUs. The paper states that it achieves performance comparable to VLAs that are 10 times larger, measured on the authors’ own simulated and real-world benchmarks. Firsthand draws no comparison from that claim.

SmolVLA is third-party research, and Firsthand is not part of it, endorsed by it or affiliated with its authors. Firsthand makes no claim that its data trained or was used to evaluate SmolVLA, and no performance claim against it. In the sources, pretraining data is community robot datasets collected from a single robot type, and egocentric human video is not described as part of the mix. What Firsthand states is custom egocentric collection for VLA-style training: synchronized egocentric video, one task instruction per episode and the action that carried it out, hardware-triggered onto one clock, to a written spec, on demand. Firsthand captures human first-person video, not robot data, so robot joint state and robot actions are Not stated, and mapping to SmolVLA’s state and action inputs is the buyer’s pipeline.

SmolVLA has two parts. The first is a pretrained vision-language model, SmolVLM-2, which uses a SigLIP encoder and a SmolLM2 language decoder. To keep inference fast, the authors skip tiling and use the global image only, with a pixel shuffle operation that limits visual tokens to 64 per frame. They also use only the first 16 layers of the language model inside the vision-language model.

The second part is an action expert of roughly 100 million parameters. It is a conditional flow matching transformer that predicts an action chunk from the vision-language features, and it interleaves cross-attention and self-attention layers rather than using only one. The paper states that the vision-language model is kept frozen and only the action expert is trained. The model card lists the training objective as flow matching and the action representation as continuous.

The authors selected a subset of 481 community datasets from Hugging Face, filtered by embodiment type, episode count, overall data quality and frame coverage. In the limitations section they put the training set at approximately 23k trajectories and say that pretraining uses datasets from a single robot type, SO100.

They report two cleanup steps. Community task annotations were noisy, with ambiguous placeholders, vague commands or no instruction at all, so an off-the-shelf vision-language model was prompted to write a short, action-oriented task description for each dataset. Camera names were also inconsistent, so each camera was manually mapped to a standard top, wrist or side view.

The paper describes an asynchronous inference stack that decouples perception and action prediction from action execution. In the usual synchronous mode, the robot executes a whole action chunk before the next prediction begins. In the asynchronous mode, prediction and control run in parallel, which the authors say allows higher control rates with chunked action generation. They compare both modes on success rate and policy speed. The figures are theirs, and this page does not repeat them.

The model card for lerobot/smolvla_base states the licence as apache-2.0, names SmolVLM2-500M-Video-Instruct as the base model and describes the intended use as a base model to fine-tune on your own use case. The LeRobot repository carries an Apache 2.0 licence badge and lists SmolVLA among its supported VLA models. The abstract says the authors release all code, pretrained models and training data. Licence terms for the training datasets themselves are Not stated in the sources read.

Fine-tuning runs through LeRobot. The model card shows installing the smolvla extra and launching lerobot-train with a dataset repository ID and the base checkpoint as the policy path.

The authors list three. Pretraining uses datasets from a single robot type, so they expect that adding multiple embodiments would help it generalize to new platforms. The roughly 23k trajectories are far fewer than typical VLA training regimes, and they note OpenVLA uses around 1 million. Finally, the model has fewer than 0.5 billion parameters, which suits consumer hardware, and the authors say they want to explore the effect of scaling it.

Write the target skills, environments and signals into a skill spec and confirm them on a Spec Pilot, a small fast first batch. This is on-demand custom collection, not a catalog download.

Where coverage is deep, the median from signed spec to first validated batch is 48 hours; where it is thin, three to four weeks. Billing is per validated hour. Firsthand exports the LeRobot v2.1 layout directly, alongside RLDS, WebDataset and HDF5.

What is SmolVLA?

A small, open vision-language-action model for robot control. Its main model has 450 million parameters and outputs continuous actions through a flow matching action expert, per the paper (arXiv 2506.01844).

How many parameters does SmolVLA have?

The paper states the main model contains 450 million parameters, with about 100 million in the action expert, and the limitations section says it has less than 0.5 billion.

Is SmolVLA open source?

The model card states the licence as apache-2.0 and the LeRobot repository carries an Apache 2.0 licence. The licence terms of the training datasets are Not stated in the sources read.

What data was SmolVLA pretrained on?

481 community datasets from Hugging Face, about 23k trajectories in total, collected from a single robot type. Egocentric human video is not described as part of the mix.

Can egocentric video fine-tune SmolVLA?

The sources do not say so. Firsthand captures first-person human video with one task instruction per episode, not robot state or robot actions, so mapping it to SmolVLA’s inputs is the buyer’s pipeline, settled in the spec.

Is Firsthand data part of SmolVLA?

No. Firsthand is not part of it, endorsed by it or affiliated with its authors, and makes no claim that its data trained or evaluated SmolVLA.