AI-MODELSAIM-001ACTIVE
VERIFIEDOPEN-WEIGHTCOMMERCIAL-OKMIT
TRUST94/100
BENCH 2024-03-12INDEXED 2024-05-01
Vision-Language-ActionManipulation Policy

OpenVLA

DEVUC Berkeley / Stanford / CMU·Embodied AI Lab

OpenVLA is an open-source 7.5B-parameter vision-language-action model trained on the Open X-Embodiment dataset, enabling generalist robot manipulation from natural language instructions and visual observations.

OpenVLA fine-tunes a Prismatic VLM backbone on 970k robot demonstrations across 70+ robot embodiments, producing action tokens directly from camera observations and language commands. It achieves state-of-the-art zero-shot generalization on the BridgeData V2 and RT-2 benchmarks and is freely available for research and commercial adaptation.

56.5%Benchmark
18Robot Compat
4Datasets
7Papers
12Related Models
MEDDeploy
SERVERHW Tier
NEAR-RTLatency
CAP::CAPABILITY_MATRIX8 AXES
IDCAPABILITYLEVELNOTES
CAP-MNPManipulation
HIGH
7-DoF Cartesian control across diverse grippers
CAP-NAVNavigation
LOW
Not primary use case; limited to short-range motion
CAP-PRCPerception
HIGH
Visual scene understanding via SigLIP-400M encoder
CAP-PLNPlanning
MED
Single-step action prediction; no explicit tree search
CAP-GRPGrasping
HIGH
Trained on 70+ robot embodiments with diverse grasp strategies
CAP-MMRMultimodal Reasoning
HIGH
Natural language grounding via Llama 2 7B backbone
CAP-TXETask Execution
HIGH
Zero-shot success on BridgeData V2 evaluation suite
CAP-TRFTransfer Learning
SOTA
Cross-embodiment zero-shot generalization on Open X-Embodiment
BENCH::PERFORMANCE_REGISTRYLAST UPDATE: 2024-03-12
BENCHMARKSCORERANKDATEDELTA
BridgeData V2 (Success Rate)56.5%#22024-03+4.2%
RT-2 Generalization Eval66.7%#32024-03+1.1%
Open X-Embodiment RT-X Eval54.2%#42024-03
LIBERO (Zero-Shot)42.0%N/A2024-06