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.

SmolVLA

Hugging Face · June 2025

Compact VLAOpen weights, Apache-2.0Also written SmolVLA-450M, lerobot/smolvla_baseTrainable on AY-Robots

Parameters: 450 M; The Hub safetensors index for lerobot/smolvla_base reports 450,046,176 parameters (446,772,624 in BF16 plus 3,273,552 in F32).
GPU memory: 0.8 GB; weights at bf16, computed. Fits an 8 GB card
Inference latency: not published; per action step
Weights: Apache-2.0; Apache-2.0 applies to the LeRobot code base.

What it is

SmolVLA is a 450M parameter vision language action model from Hugging Face, built on the SmolVLM2-500M video backbone with a flow matching action expert that predicts chunks of 50 actions. It cuts cost in two ways: only the first half of the language model layers are executed, and visual input is capped at 64 tokens per frame. Pretraining used exclusively openly licensed LeRobot community datasets (481 datasets, 22.9K episodes, 10.6M frames), one to two orders of magnitude less data than pi0 or OpenVLA consumed. It reaches 87.3 percent on the LIBERO average, ahead of pi0 at 86.0, but only 57.3 percent on Meta-World, where the harder tiers expose its ceiling. The practical draw is that it fine-tunes on a single consumer GPU and ships inside LeRobot alongside the SO-100 and SO-101 arms it was actually validated on.

Architecture

Backbone: SmolVLM2-500M-Video-Instruct (SigLIP vision encoder plus a SmolLM2 language decoder). Only the first half of the language model layers are executed, and visual input is capped at 64 tokens per frame.
Action head: flow matching action expert, 10 integration steps at inference, with interleaved cross attention and causal self attention blocks
Action chunk: 50 steps per forward pass
Parameters: The Hub safetensors index for lerobot/smolvla_base reports 450,046,176 parameters (446,772,624 in BF16 plus 3,273,552 in F32). Roughly 100M of that is the flow matching action expert, whose hidden dimension is 0.75x the VLM hidden size. The paper additionally ablates 0.24B and 2.25B variants (LIBERO averages 82.75 and 88.75 percent), but only the 0.45B checkpoint is released.
Pretraining data: Community contributed LeRobot datasets. The paper reports 481 datasets, 22.9K episodes and 10.6M frames, while the official Hugging Face blog post says 487 curated datasets, so the two official sources disagree by six datasets. The corpus is dominated by SO-100 class arm data recorded at 30 FPS and is at least an order of magnitude smaller than the corpora behind pi0 or OpenVLA. Pretraining ran on 4 GPUs at batch size 256 for 200,000 steps.
Embodiments: SO-100, SO-101, LeKiwi, LIBERO (simulation), Meta-World (simulation)

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 450 M 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	1.7 GB	Training master weights, rarely used for inference
bf16 or fp16	0.8 GB	The usual way these checkpoints are served
int8	0.4 GB	Quantised, expect some loss of precision on fine motions
int4	0.2 GB	Aggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 8 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.

No source publishes a VRAM figure or a per step latency. The paper states the model can be trained on a single GPU and deployed on consumer grade GPUs or even CPUs, and the official Hugging Face blog states it is small enough to run on CPU, train on a single consumer GPU, or even a MacBook. Pretraining itself used 4 GPUs at batch size 256.

The authors publish no memory or latency figure for this model. Everything above is computed from the parameter count.

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 spatial, object, goal and long
87.3%
success rateAuthors' own evaluation. Same table reports pi0 (3.3B) at 86.0 and OpenVLA (7B) at 76.5.source
LIBERO long
71%
success rateWeakest LIBERO suite, below pi0 at 73.source
Meta-World average over easy, medium, hard and very hard
57.3%
success rateAuthors' own evaluation. Per tier: easy 82.5, medium 41.8, hard 45.0, very hard 60.0. Same table reports pi0 (3.5B) at 47.9 and TinyVLA at 31.6.source

Real world

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

SO-100 real robot 3 tasks multi-task (pick-place, stacking, sorting)
78.3%
success rateAuthors' own real hardware evaluation. Community pretraining lifted this from 51.7 to 78.3 percent. pi0 scored 61.7 in the same setup.source
SO-101 real robot Pick-Place-Lego, out of distribution
50%
success rateAuthors' own evaluation. In distribution on the same task is 90 percent, so the drop under distribution shift is 40 points.source

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Validated by the authors on SO-100 and SO-101 low cost arms. On SO-100 it averaged 78.3 percent across pick and place, stacking and sorting in a multi-task setting, against 61.7 percent for pi0. On SO-101 it reached 90 percent in distribution and 50 percent out of distribution on a Lego pick and place task. Asynchronous inference, where action prediction is decoupled from execution, cut average task time from 13.75 to 9.70 seconds and raised throughput from 9 to 19 completed cubes in a fixed 60 second window, at the cost of success rate falling from 78.3 to 73.3 percent.

Fine tuning it yourself

Fine-tuned with the lerobot-train entry point on any dataset in LeRobot format, starting from the policy path lerobot/smolvla_base. The model card documents dataset repo id, batch size, step count, bfloat16 training and a policy.chunk_size flag. Neither the paper nor the model card states a minimum number of demonstrations.

This is one of the models the AY-Robots training pool can fine tune for you on a dataset recorded with your own arm. The platform page carries the defaults, GPU tier and inference latency measured in that pool.

Platform page for SmolVLA

Where it helps, where it does not

Strengths

Beats pi0 (3.3B) on the LIBERO average with roughly 7x fewer parameters, 87.3 against 86.0 percent.
Asynchronous inference roughly doubles task throughput in a fixed time window (19 against 9 cubes) and cuts average task time by about 30 percent.
Pretrained only on openly licensed community datasets, so the entire pretraining corpus is inspectable and reproducible.
Ships inside LeRobot, so fine-tuning, teleoperation and the SO-100 or SO-101 hardware path all use one toolchain.
Smallest credible entry point of this group: 450M parameters, single consumer GPU training.

Limits

Meta-World average of 57.3 percent means it fails roughly half of the medium and harder tasks.
LIBERO-Long at 71 percent is its weakest suite and trails pi0 at 73.
Pretraining is dominated by SO-100 class low cost arms, so transfer to industrial arms or different kinematics is not demonstrated.
Asynchronous inference trades accuracy for throughput: success drops from 78.3 to 73.3 percent in the authors' own comparison.
No published VRAM or per step latency figure, so edge deployment has to be measured rather than looked up.
The paper and the official blog post disagree on the pretraining corpus size (481 against 487 datasets).
The released checkpoint is the 0.45B variant only. The stronger 2.25B ablation is not published.

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