OpenVLA
The de-facto open-source VLA baseline — 7B parameters, MIT license, and 1.52M monthly HuggingFace downloads, beating Google's 55B RT-2-X by 16.5 points.
1. Core Product / Service
OpenVLA is an open-source 7-billion-parameter Vision-Language-Action (VLA) model for robotic manipulation, developed by researchers from Stanford, UC Berkeley, Google DeepMind, Toyota Research Institute (TRI), and MIT [1]. Released in June 2024 under the MIT license, it was the first credibly open competitor to Google's closed RT-2 series.
Key architecture decisions:
- Dual visual encoders: DINOv2 (spatial reasoning) + SigLIP (semantic understanding) — a novel approach over single-encoder competitors at the time [1].
- Base LLM: Llama 2 7B, fine-tuned on ~970K robot manipulation episodes from the Open X-Embodiment dataset spanning 70+ datasets and 22 robot embodiments.
- Discrete action tokenization: 256 bins per action dimension, mapped to unused Llama 2 tokens.
- Fine-tuning access: LoRA/QLoRA support enables adaptation on consumer GPUs with as little as 8 GB VRAM.
The February 2025 OpenVLA-OFT update replaced discrete action tokens with a continuous regression head and added action chunking, achieving a 26× throughput improvement and 97.1% LIBERO success rate — up from 76.5% [1]. A PyTorch training pipeline released March 2025 supports scaling from 10B to 340B parameters.
On HuggingFace, OpenVLA-7B records ~1.52M monthly downloads — 36× more than the #2 model (GR00T-N1.7-3B at 41.4K), making it by far the most-used robot model on the platform [2].
2. Target Users & Pain Points
- Academic robotics researchers: Need a strong, open baseline to build upon — OpenVLA fills the gap that Google's closed RT-2 left. Most subsequent VLA papers benchmark against or build on top of OpenVLA.
- Robotics startups: Get a production-quality VLA model without paying for API access or building from scratch. The MIT license enables commercial use.
- Edge deployment engineers: The LoRA/QLoRA + quantization stack allows running VLA models on edge hardware — MiniVLA and other derivatives target this use case.
The core pain point: building a VLA from scratch requires GPU clusters and massive robot datasets that most labs can't access. OpenVLA democratizes the starting point.
3. Competitive Landscape
| Model | Type | Open Source | Scale |
|---|---|---|---|
| Google RT-2 / RT-2-X | Closed VLA | No | 55B params |
| GR00T N1 → N1.7 | Open VLA + WBC | Yes (weights, not code) | 3B |
| PI π0.5 | Closed VLA | No (HF hosted) | 3–4B |
| G0.5 (Galaxea) | Open VLA (unified autoregressive) | Yes | ~2B |
OpenVLA's differentiation is its combination of open-source completeness (code + weights + training pipeline, MIT license) and academic neutrality — it's not tied to any single robot manufacturer or cloud provider, making it the default baseline that subsequent models benchmark against.
4. Unique Observations
- 36× the downloads of GR00T. At 1.52M monthly HuggingFace downloads vs GR00T-N1.7-3B's 41.4K, OpenVLA dominates robot model usage by an order of magnitude [2]. This suggests the developer community strongly prefers an MIT-licensed, truly open model over a corporate-open model tied to a specific hardware ecosystem.
- The academic-commercial flywheel. Because OpenVLA is the default baseline, every new VLA paper benchmarks against it → more citations and community trust → more downloads → more derivative work → OpenVLA remains the baseline. This is the same dynamic that made BERT and ResNet dominant in their respective eras.
- From lab to production. OpenVLA-OFT's 26× speedup and quantization support are bending the curve toward real-time robot inference on affordable hardware. If 30+ Hz VLA inference on a Jetson becomes routine, the barrier between research demo and deployed robot drops dramatically.
5. Financials / Funding
OpenVLA is a research project, not a commercial entity — no separate funding rounds. Development has been supported by:
- Stanford University, UC Berkeley, Google DeepMind, TRI, and MIT — the contributing institutions fund the researchers.
- The project is MIT-licensed and community-maintained via GitHub (github.com/openvla/openvla) and HuggingFace.
6. People & Relationships
- Key authors: Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, Chelsea Finn
- Hosting: HuggingFace (openvla/openvla-7b), GitHub (openvla/openvla)
- Competitors / benchmarks against: Google RT-2, nvidia-groot, physical-intelligence, Galaxea G0.5
- Part of: Open X-Embodiment collaboration, LeRobot ecosystem
OpenVLA
7B 开源 VLA，用 OXE 筛选出的约 97 万条轨迹训练
- 数据类型
- 真机遥操演示
- 规模口径
- 约 970K 条轨迹，源自 OXE（70+ 数据集、200 万+ 轨迹）
- License
- 代码 MIT；权重受 Llama 2 Community License 约束
数据配方
骨干为 Llama 2 7B，视觉端融合 DINOv2 与 SigLIP [1]。数据取自 open-x-embodiment（70+ 数据集、200 万+ 轨迹），仅保留至少一路第三人称相机、单臂末端执行器控制的操作数据，沿用 Octo 的混合权重（降低低多样性数据、提高任务与场景多样性数据），最终约 97 万条 [2]。droid 以 10% 权重加入，因动作 token 准确率偏低，训练末三分之一移除并重分配权重 [2]。数据先转为 rlds 格式，BridgeData V2 需用官方新版替换 OXE 旧版（见 bridgedata-v2），自有数据通过 transforms 与 dataset config 注册标准化 [3]。"22 本体"为 OXE 原论文口径，OpenVLA 论文该节未复述 ⚠️。
结果与证据
29 个任务上绝对成功率比 RT-2-X（55B）高 16.5 个百分点，参数少 7 倍 [1]。Franka 微调中是唯一在所有测试任务达到至少 50% 成功率的方法；LoRA 仅训练约 1.4% 参数即可匹配全量微调 [4]。
已知缺陷与口径
仅支持单图输入，不含本体状态与历史；无 action chunking，高频（如 ALOHA 50Hz）受限，成功率通常低于 90% [2]。数据中的停顿动作会致推理卡住，建议 5–10Hz 采集 [3]。语义泛化弱于 RT-2-X；训练耗 64 张 A100、15 天 [4]。许可方面 MIT 仅覆盖代码，权重继承 Llama 2 限制 [3]。