How is pricing calculated?

Pricing is based on the hourly rate of operators in your selected region. You only pay for the time operators spend controlling your robots.

What robots are supported?

We support 9+ robot arms including SO-100, Koch, Franka FR3, Franka Panda, WidowX 250, ViperX 300, ViperX 300s, and ALOHA.

What is the latency for teleoperation?

Our platform provides ultra-low latency of 30-70ms worldwide, enabling real-time responsive robot control.

OpenVLA

Stanford University, USA · June 2024

Stanford University, UC Berkeley, Toyota Research Institute, Google DeepMind, Physical Intelligence, MIT

Foundation VLAOpen weights, MITAlso written OpenVLA-7B, openvla/openvla-7b

Parameters: 7.5 B; The Hugging Face safetensors index for openvla/openvla-7b reports 7,541,237,184 BF16 parameters.
GPU memory: 16.8 GB; reported by the authors
Inference latency: 239.6 ms; 239.6 ms per action on an NVIDIA A100, averaged over 100 queries with one 224 by 224 px image plus a LIBERO instruction,...
Weights: MIT; MIT for the code and on the Hugging Face model card.

What it is

OpenVLA is a 7B vision-language-action model that fine-tunes a Llama 2 7B VLM with a fused DINOv2 and SigLIP vision encoder on 970k Open X-Embodiment episodes. It emits 7 discrete action tokens per timestep, one per end-effector dimension, and controls a robot at roughly 6 Hz on an RTX 4090. It was the first VLA to ship weights, code and data mixture openly, and it remains the reference baseline for the field. Its two structural weaknesses are the single-image observation space and the absence of action chunking, which together rule out wrist cameras, proprioception and high-frequency bimanual control. As of August 2026 there is no successor release: the arXiv paper is still at v3 from September 2024 and the repository was last pushed in March 2025.

Architecture

Backbone: Prismatic-style VLM with a Llama 2 7B language model and a fused visual encoder combining DINOv2 and SigLIP features
Action head: Autoregressive discrete action tokens, 256 bins per dimension, 7 tokens decoded sequentially per timestep
Action chunk: 1 step per forward pass
Control rate: 6 Hz
Parameters: The Hugging Face safetensors index for openvla/openvla-7b reports 7,541,237,184 BF16 parameters. The paper markets the model as 7B. Note a caveat from footnote 4 of the paper: the parameter-efficient fine-tuning study (Section 5.3) and the quantization study (Section 5.4) use a smaller SigLIP-only variant with 7,188.1M parameters, not the released DINOv2 plus SigLIP checkpoint.
Pretraining data: 970k real-world robot demonstration episodes curated from Open X-Embodiment (the authors call the mixture Open-X Magic Soup++). Pretraining ran on a cluster of 64 A100 GPUs for 14 days, 21,500 A100-hours in total, at batch size 2048.
Embodiments: WidowX (BridgeData V2), Google robot (RT-1 and RT-2 mobile manipulator), Franka Emika Panda (Franka-Tabletop, via fine-tuning)

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Computed from 7.5 B parameters. Weights only, so treat it as a floor: activations, image encoder intermediates and the CUDA context come on top, in practice 30 to 50 percent more for inference.
Precision	Weights	Note
fp32	28 GB	Training master weights, rarely used for inference
bf16 or fp16	14 GB	The usual way these checkpoints are served
int8	7.0 GB	Quantised, expect some loss of precision on fine motions
int4	3.5 GB	Aggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 24 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.

Published figures. These come from the model authors, not from this site.
Inference memory, reported	16.8 GB
Fine tuning memory, reported	59.7 GB
Inference latency, reported	239.6 ms on 239.6 ms per action on an NVIDIA A100, averaged over 100 queries with one 224 by 224 px image plus a LIBERO instruction, measured in the OpenVLA-OFT paper (Table II). The OpenVLA paper itself reports roughly 6 Hz on a single NVIDIA RTX 4090 without compilation or speculative decoding, 3 Hz with int4 and only 1.2 Hz with int8 on an A5000. source
Control rate	6 Hz
Checkpoint download	15.09 GB

The GitHub README states LoRA fine-tuning works on any GPU with at least about 27 GB of memory, and that the reference configuration (batch size 16, no gradient accumulation) needs about 72 GB and was run on a single A100 80 GB. Full fine-tuning is only recommended with a full node of 8 A100 GPUs. For inference the paper reports 7.0 GB with int4 quantization at no measured accuracy loss.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

Reproducible benchmark suites. Everyone runs the same episodes, so the numbers can be compared inside a suite.

LIBERO average over LIBERO-Spatial, Object, Goal and Long, fine-tuned per suite, 3 seeds with 500 trials each
76.5%
success rateReported by the OpenVLA authors, Table 12 (76.5 plus or minus 0.6). Fine-tuned Octo reaches 75.1% and Diffusion Policy from scratch 72.4% under the same protocol.source
LIBERO LIBERO-Long, fine-tuned, 3 seeds with 500 trials
53.7%
success rateWeakest of the four suites and the number that OpenVLA-OFT later raised to 94.5%. Reported by the OpenVLA authors, Table 12.source

Real world

Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.

Google robot (real RT-1 mobile manipulator) 12 tasks, 5 trials each, 60 rollouts
85%
mean success rateReported by the OpenVLA authors, Table 6 (85.0 plus or minus 4.6). RT-2-X reaches 78.3%; the authors bold both because the error bars overlap.source
BridgeData V2 (real WidowX) 8-task quantization subset, 80 rollouts per precision
71.9%
success rate with int4 weightsReported by the OpenVLA authors, Tables 2 and 5. bfloat16 reaches 71.3% at 16.8 GB, int8 only 58.1% at 10.2 GB. The authors attribute the int8 drop to speed, not precision: int8 runs at 1.2 Hz on their A5000 while the data was collected with a 5 Hz controller.source
BridgeData V2 (real WidowX) full 17-task evaluation suite, 170 rollouts per method
70.6%
mean success rateReported by the OpenVLA authors, Table 4 (70.6 plus or minus 3.2). RT-2-X reaches 50.6%, Octo 20.0% and RT-1-X 18.5% on the identical suite. Partial credit of 0.5 is possible on some tasks.source
Franka-Tabletop (real Franka Emika Panda) selected tasks, 33 rollouts per fine-tuning strategy
68.2%
success rate with LoRA rank 32Reported by the OpenVLA authors, Table 1. Full fine-tuning reaches 69.7% while training 7,188.1M parameters; LoRA rank 32 trains only 97.6M parameters, 1.4% of the model.source

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Evaluated on two real platforms without fine-tuning: WidowX from BridgeData V2 at 70.6% over 170 rollouts and the Google robot at 85.0% over 60 rollouts. Fine-tuned on a real Franka Emika Panda (Franka-Tabletop) it reaches 68.2% with LoRA rank 32 over 33 rollouts. All evaluations were run as A/B comparisons with identical initial object and robot states.

Fine tuning it yourself

LoRA applied to all linear layers, recommended rank r = 32 (rank has negligible effect on performance). Training data must be in RLDS format; the repo ships LIBERO RLDS conversions of about 10 GB and documents the BridgeData V2 download at 124 GB. One new task takes 10 to 15 hours on a single A100, an 8x compute reduction versus full fine-tuning. Full fine-tuning is only advised on a node of 8 A100 GPUs.

Where it helps, where it does not

Strengths

Fully open release of weights, training code and the data mixture, which is why it became the standard baseline that later VLAs compare against
Strong language grounding in cluttered multi-object scenes, beating the closed 55B RT-2-X by 20.0 points on the BridgeData V2 suite despite being roughly 7 times smaller
LoRA rank 32 matches full fine-tuning on Franka-Tabletop while training 1.4% of the parameters, and one new task takes 10 to 15 hours on a single A100
int4 quantization cuts inference memory to 7.0 GB with no measured accuracy loss on BridgeData V2
Hugging Face AutoModel integration, so loading and serving follow standard LLM tooling

Limits

Accepts only a single third-person image. No wrist camera, no proprioceptive input and no observation history, which the authors name as the first limitation in Section 6.
No action chunking. Every timestep costs 7 sequential decoder passes, giving roughly 3 to 6 Hz, far below the 25 to 50 Hz that bimanual platforms such as ALOHA need. The authors call this out explicitly.
Reliability tops out below 90% success on the tested tasks, so it is not dependable for unattended operation.
Cannot zero-shot control embodiments that are absent from the pretraining mixture; the model card states this directly.
int8 quantization is a trap: it saves memory but drops control to 1.2 Hz on an A5000 and costs 13 points of success rate. int4 is the better quantization choice.
LoRA fine-tuning at the documented reference batch size needs about 72 GB, so a consumer 24 GB card requires gradient accumulation and much longer runs.

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-08-11.

Try a policy on a real arm

A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.

Download the client Drive a real arm now