LeRobot SmolVLA

Σχετικά με Hugging Face (LeRobot team)

Ιδρύθηκε 2016 · New York, USA / Paris, France

SmolVLA is the flagship Vision-Language-Action model of Hugging Face's LeRobot project, an open-source robotics framework that brings the Transformers / Datasets philosophy to physical-AI research. SmolVLA was released in mid-2025 as a deliberately compact 450M-parameter VLA designed to be trainable and runnable on consumer hardware while still benefiting from community-scale pretraining. It is trained on 487 publicly contributed LeRobot community datasets - teleoperation episodes uploaded by hobbyists, university labs and small robotics companies - making it the first community-data-driven open VLA. The release includes pretraining and fine-tuning code, model checkpoints under Apache-2.0, and a tightly integrated stack with the LeRobot framework, hf-hub-hosted datasets, and the SO-100 / SO-ARM-100 low-cost robot arms.

Αρχιτεκτονική

Compact Vision-Language-Action transformer (action-chunk regression)

SmolVLA is a 450M-parameter transformer that combines a SmolVLM-style vision-language encoder with an action expert that regresses continuous action chunks. The vision-language tower is initialised from the open SmolVLM family (compact VLMs released by Hugging Face) and is responsible for fusing multi-view RGB observations with the natural-language instruction; a smaller action-prediction head consumes the resulting tokens together with proprioception and outputs a short chunk of continuous joint or end-effector actions. The model is pretrained on 487 LeRobot-format community datasets, covering single-arm, dual-arm and mobile-base setups, with a strong tilt toward the popular SO-100 and Koch low-cost teleoperation arms. Post-pretraining, users fine-tune on their own LeRobot recording for a specific robot and task. The whole stack is designed to run pretraining on a few H100s and fine-tuning on a single consumer GPU.

Παράμετροι

450M

Δυνατότητες

• Compact 450M open VLA pretrained on community data
• Trained on 487 LeRobot community datasets
• SmolVLM-style vision-language tower + action expert
• Continuous action-chunk regression
• Runs fine-tuning on a single consumer GPU
• Tight integration with LeRobot framework on Hugging Face
• Apache-2.0 licence on weights and code
• Strong baseline for SO-100 and Koch low-cost arms
• Best for: hobbyists, educators, low-cost robot research.

Εκπαίδευση & άδεια

487 publicly contributed LeRobot-format community datasets hosted on the Hugging Face Hub, dominated by teleoperation episodes from low-cost arms (SO-100, Koch) but also including dual-arm and mobile setups. Total scale on the order of millions of frames.

Άδεια: Apache-2.0 - fully open weights, code, and datasets (where contributors used compatible licences). Designed for both research and commercial use.

Δοκιμές ασφάλειας: Research / hobbyist artifact - no formal red-teaming. Safety in deployment relies on low-torque hardware (SO-100, etc.) and user-supplied workspace constraints.