SmolVLA: Efficient Vision-Language-Action Model trained on Lerobot Community Data

Today, we introduce SmolVLA , a compact (450M), open-source Vision-Language-Action model for robotics that runs on consumer hardware.

Over the past few years, Transformers have driven remarkable progress in AI, from language models capable of human-like reasoning to multimodal systems that understand both images and text. However, in real-world robotics, advancements have been much slower. Robots still struggle to generalize across diverse objects, environments, and tasks. This limited progress stems from a lack of high-quality, diverse data and the absence of models that can reason and act like humans in the physical world .

In response to these challenges, the field has recently turned to vision-language-action (VLA) models , which aim to unify perception, language understanding, and action prediction within a single architecture. VLAs typically take as input raw visual observations and natural language instructions, and output corresponding robot actions. While promising, much of the recent progress in VLAs remains locked behind proprietary models trained on large-scale private datasets, often requiring costly hardware setups and extensive engineering resources. As a result, the broader robotics research community faces significant barriers in reproducing and building upon these models.

SmolVLA addresses this gap by offering an open-source, compact, and efficient VLA model that can be trained on consumer-grade hardware using only publicly available datasets . By releasing not only model weights but also using very affordable open-source hardware, SmolVLA aims to democratize access to vision-language-action models and accelerate research toward generalist robotic agents.

​​If you'd like to build from the architecture (pretrained VLM + action expert) rather than a pretrained checkpoint:

SmolVLA is not only a lightweight yet capable model, but also a method for training and evaluating generalist robotics policies. In this section, we introduce the model architecture behind SmolVLA and the asynchronous inference setup used for evaluation, which has proven to be more adaptable and capable of faster recovery.

SmolVLA consists of two core components: a Vision-Language Model (VLM) that processes multimodal inputs and an action expert that outputs robot control commands. Below, we share the details of the main components of SmolVLA architecture and the Asynchronous Inference. More details can be found in our technical report .

Figure 1: Comparison of SmolVLA across task variations. From left to right: (1) asynchronous pick-place cube counting, (2) synchronous pick-place cube counting, (3) pick-place cube counting under perturbations, and (4) generalization on pick-and-place of the lego block with real-world SO101.

In SmolVLA, CA ensures that actions are well-conditioned on perception and instructions, while SA improves temporal smoothness —especially critical for real-world control, where jittery predictions can result in unsafe or unstable behavior.

SmolVLA is our first step toward that vision : It is pretrained on a curated mix of publicly available, community-contributed datasets designed to reflect real-world variation. Rather than optimizing for dataset size alone, we focus on diversity: a range of behaviors, camera viewpoints, and embodiments that promote transfer and generalization.

Inspired by the training paradigms of Large Language Models (LLMs), SmolVLA goes through a pretraining phase on general manipulation data, followed by task-specific post-training. Architecturally, it combines Transformers with flow-matching decoders , and is optimized for speed and low-latency inference with the following design choices:

Despite using fewer than 30k training episodes—an order of magnitude less than other VLAs—SmolVLA matches or exceeds the performance of much larger models, both in simulation and the real world.

To make real-time robotics easier to use, we introduce an asynchronous inference stack. This technology separates how robots perform actions from how they understand what they see and hear. Because of this separation, robots can respond more quickly in fast-changing environments.