Vision-Language-Action Robot Policies
Robot policies that start from a pretrained vision-language model and output motor actions from camera images and a language instruction.
Tier Grades
Tier A: Frontier AI labs
feasible
Physical Intelligence, NVIDIA and Google DeepMind train generalist VLAs on large multi-robot mixtures such as OpenVLA's 970k demonstrations.
Tier B: Indonesian AI labs
stretch
Pretraining a VLA on 970k demonstrations is out of reach on the evidence, but open 7B weights can be fine-tuned with low-rank adaptation on consumer GPUs.
Tier C: Institution units such as ITB
stretch
A lab with one robot arm can fine-tune open OpenVLA or pi0 weights on its own demonstrations, but collecting robot data needs hardware.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
feasible
Fine-tuning an open VLA on consumer GPUs for one site task, such as retail picking, fits a proof of concept.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeOpenVLA: 7B parameters built on Llama 2 with DINOv2 and SigLIP features; fine-tunable on consumer GPUs with low-rank adaptation; pretraining compute not stated in the abstract.
- DataOpenVLA: 970k real-world robot demonstrations; pi0: a large mixture from single-arm, dual-arm and mobile manipulators; GR00T N1: real-robot trajectories, human videos and synthetic data.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 18 authors on the OpenVLA paper (arXiv 2406.09246); 24 on pi0; 43 on GR00T N1.
- Time horizonMulti-year.
- Shelf-life signalsOpenVLA (7B) beat the closed RT-2-X (55B) by 16.5 points across 29 tasks, so open models are overtaking closed ones within a year., SmolVLA reports performance comparable to VLAs 10x larger, so small open VLAs date quickly.
- Pipeline roleTier A releases open VLA weights; Tier B and Tier C fine-tune them on local robots and tasks; Tier D carries a fine-tuned policy into a site proof of value.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
Observed actors
Stanford University
UC Berkeley
Physical Intelligence
NVIDIA
Google DeepMind
Figure AI
AgiBot (Zhiyuan)
Galbot
Unitree Robotics
Hugging Face
BAAI
Allen AILinked models
Gemini Robotics 1.5robot policyexemplarGR00T N1robot policyexemplarπ0robot policyexemplarOpenVLArobot policyexemplar
FLUX.3 Action Baserobot policyproduced byHelix 2.5robot policyproduced byGemini Robotics 2robot policyproduced byGemini Robotics 2 VLArobot policyproduced byGemini Robotics On-Device 2robot policyproduced by
MiniCPM RobotManip and RobotTrackrobot policyproduced byMolmoAct2robot policyproduced byπ0.7robot policyproduced byGO-2robot policyproduced byGR00T N1.7robot policyproduced byUnifoLM-VLA-0robot policyproduced byHelix 02robot policyproduced byGR00T N1.6robot policyproduced byπ*0.6robot policyproduced by
X-VLArobot policyproduced byRoboBrain 2.0robot policyproduced byGemini Robotics On-Devicerobot policyproduced bySmolVLArobot policyproduced byGraspVLArobot policyproduced byπ0.5robot policyproduced by+19
GO-1robot policyproduced byGemini Roboticsrobot policyproduced byOpenVLA-OFTrobot policyproduced byHelixrobot policyproduced byπ0-FASTrobot policyproduced byRDT-1Brobot policyproduced byRT-2-Xrobot policyproduced byGOrobot policyproduced byHelixrobot policyproduced byGroceryVLArobot policyproduced byTrackVLArobot policyproduced byGemini Roboticsrobot policyproduced byRT-2robot policyproduced byGR00Trobot policyproduced byGR00T N2robot policyproduced byπ (pi) modelsrobot policyproduced byOpenVLArobot policyproduced byQwen2.5-VLmultimodal LLMused byUnifoLM-VLM-Basemultimodal LLMused byLinks
R&D TopicVision-Language-Action Robot Policies
ModelsOpenVLAStanford University, robot policyπ0Physical Intelligence, robot policyGR00T N1NVIDIA, robot policyGemini Robotics 1.5Google DeepMind, robot policy+39 more
OpenVLAStanford University, robot policyπ0Physical Intelligence, robot policyGR00T N1NVIDIA, robot policyGemini Robotics 1.5Google DeepMind, robot policy+39 more
Models 43
Gemini RoboticsGoogle DeepMind, robot policyproduced byHelix 02Figure AI, robot policyproduced byOpenVLAStanford University, UC Berkeley, Toyota Research Institute (TRI), robot policyexemplarπ (pi) modelsPhysical Intelligence, robot policyproduced byGemini RoboticsGoogle DeepMind, robot policyproduced byGemini Robotics 2Google DeepMind, robot policyproduced byGR00TNVIDIA, robot policyproduced byGR00T N1.6NVIDIA, robot policyproduced byGR00T N1.7NVIDIA, robot policyproduced byπ*0.6Physical Intelligence, robot policyproduced byGemini Robotics 1.5Google DeepMind, robot policyexemplarHelixFigure AI, robot policyproduced by+31
HelixFigure AI, robot policyproduced byUnifoLM-VLM-BaseUnitree Robotics, multimodal LLMused byπ0Physical Intelligence, robot policyexemplarπ0.5Physical Intelligence, robot policyproduced byGemini Robotics 2 VLAGoogle DeepMind, robot policyproduced byGemini Robotics On-DeviceGoogle DeepMind, robot policyproduced byGemini Robotics On-Device 2Google DeepMind, robot policyproduced byGOAgiBot (Zhiyuan), robot policyproduced byGO-1AgiBot (Zhiyuan), robot policyproduced byGO-2AgiBot (Zhiyuan), robot policyproduced byGR00T N1NVIDIA, robot policyexemplarGR00T N2NVIDIA, robot policyproduced byHelix 2.5Figure AI, robot policyproduced byOpenVLAStanford University, robot policyproduced byOpenVLA-OFTStanford University, robot policyproduced byQwen2.5-VLmultimodal LLMused byπ0-FASTPhysical Intelligence, robot policyproduced byπ0.7Physical Intelligence, robot policyproduced byGraspVLAGalbot, Peking University, robot policyproduced byGroceryVLAGalbot, robot policyproduced byRT-2Google DeepMind, robot policyproduced byRT-2-XGoogle DeepMind, robot policyproduced byUnifoLM-VLA-0Unitree Robotics, robot policyproduced byX-VLATsinghua University, Shanghai AI Lab, robot policyproduced byFLUX.3 Action BaseBlack Forest Labs, robot policyproduced byMiniCPM RobotManip and RobotTrackOpenBMB (Tsinghua NLP Lab and ModelBest), robot policyproduced byMolmoAct2Allen AI, robot policyproduced byRDT-1BTsinghua University, robot policyproduced byRoboBrain 2.0BAAI, robot policyproduced bySmolVLAHugging Face, robot policyproduced byTrackVLAGalbot, robot policyproduced byPapers
- RT-1: Robotics Transformer for Real-World Control at ScaleAnthony Brohan et al., arXiv preprint, 2022 founding
- Gemini Robotics: Bringing AI into the Physical WorldGemini Robotics Team et al., arXiv preprint, 2025 state of the art
- Octo: An Open-Source Generalist Robot PolicyOcto Model Team et al., arXiv preprint, 2024 state of the art
- OpenVLA: An Open-Source Vision-Language-Action ModelMoo Jin Kim et al., arXiv preprint, 2024 state of the art
- A Survey on Vision-Language-Action Models for Embodied AIYueen Ma et al., IEEE Transactions on Neural Networks and Learning Systems (Early Access), 2026, 2024 survey
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- OpenVLA: An Open-Source Vision-Language-Action Model arxiv.org https://arxiv.org/abs/2406.09246
- arxiv.org/abs/2410.24164 https://arxiv.org/abs/2410.24164
- GR00T N1: An Open Foundation Model for Generalist Humanoid Robots arxiv.org https://arxiv.org/abs/2503.14734
- arxiv.org/abs/2506.01844 https://arxiv.org/abs/2506.01844
- arxiv.org/abs/2507.02029 https://arxiv.org/abs/2507.02029
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Vision-Language-Action Robot Policies", https://aurion.semai.id/topics/vision-language-action-robot-policies/, records of October 2026.