Robot learning policy models, compared
Five models can be trained on a dataset you record with your own arm. They differ in what they cost to train, how fast they run on hardware, and how much data they need before they do anything useful. This page puts the numbers side by side.
- 5 supported models
- LeRobot v2.1 datasets
- GPUs rented by the hour
All five models at a glance
The GPU tier is what the training pool requests, not a recommendation you can override. A model on the 80 GB tier cannot be fine-tuned on a consumer card, and inference latency is measured per action step in that same pool.
| Model | Vendor | Parameters | GPU tier | Inference | Default steps | Min. episodes |
|---|---|---|---|---|---|---|
| GR00T N1.7 | NVIDIA | about 3 billion | A100 80 GB or H100 80 GB | 152 ms per step | 20,000 | 50 |
| Pi0.5 | Physical Intelligence | about 3 billion | A100 80 GB or H100 80 GB | 485 ms per step | 30,000 | 50 |
| SmolVLA | Hugging Face | about 450 million | RTX 4090 or any card with 24 GB | 245 ms per step | 20,000 | 30 |
| ACT | Stanford (ALOHA) | about 80 million | RTX 4090 or any card with 24 GB | 20 ms per step | 100,000 | 50 |
| GR00T N1.5 | NVIDIA | about 3 billion | A100 80 GB or H100 80 GB | 165 ms per step | 2,000 | 50 |
Default steps and the minimum episode count are starting points the training form fills in for you, not limits. Every trainer counts steps, none of them counts epochs.
Pick a model
Each page covers what the model actually is, the dataset it expects, the knobs the training form exposes for it, and the failure modes people hit first.
The strongest starting point when you have a proper dataset and want the highest success rate.
- GPU
- A100 80 GB or H100 80 GB
- Inference
- 152 ms per action step
- Data
- from 50 episodes
- Format
- LeRobot v2.1
Typical run 3 to 6 hours, about 4 to 12 USD
Highest precision on contact-rich tasks, at roughly three times the inference latency of GR00T.
- GPU
- A100 80 GB or H100 80 GB
- Inference
- 485 ms per action step
- Data
- from 50 episodes
- Format
- LeRobot v2.1
Typical run 3 to 6 hours, about 4 to 12 USD
The default first run: cheap enough to iterate on, good enough to tell you whether your data is any good.
- GPU
- RTX 4090 or any card with 24 GB
- Inference
- 245 ms per action step
- Data
- from 30 episodes
- Format
- LeRobot v2.1
Typical run 2 to 5 hours, about 1 to 3 USD
No pretraining, no language, no foundation model. It learns exactly your task and nothing else, very fast.
- GPU
- RTX 4090 or any card with 24 GB
- Inference
- 20 ms per action step
- Data
- from 50 episodes
- Format
- LeRobot v2.1
Typical run 2 to 5 hours, about 1 to 3 USD
Kept available for reproducing older runs. New projects should use N1.7.
- GPU
- A100 80 GB or H100 80 GB
- Inference
- 165 ms per action step
- Data
- from 50 episodes
- Format
- LeRobot v2.1
Typical run 3 to 6 hours, about 4 to 12 USD
Head to head
Six pairings that come up in practice, each one answering a single question with the same dataset on both sides.
Two 3 B foundation models on the same card: highest success rate against the smoothest contact behavior.
An 80 GB fine-tune against a run you can afford to repeat every evening.
The two models that train on consumer hardware: language conditioning against raw speed.
Flow matching on an A100 against a 450 M model on a 24 GB card.
Pretrained generalist against a model that learns only your task, at 20 ms per step.
What changed between the two GR00T releases, and why new projects should not start on N1.5.
Try the arm before you pick a model
A real SO-100 is online and free to drive in the browser. No signup, no client, no hardware of your own.