Train a policy on your arm
Every supported model on every supported arm, with the defaults the training backend really sends. Pick the row for your model and the column for your hardware.
- 17 guides
- 5 models, 4 arms
- LeRobot v2.1 datasets
Find your combination
Rows are the models the training backend supports, columns are the arms the client can drive. A cell links to the guide for that exact pair. Where a cell has no link, that combination is not written up yet: the closest guide is the same model on another arm, since only the driver and the wiring change. If a whole row is empty, the model reference behind the model name has the parameters.
| Model | SO-100Full support | SO-101Full support | Koch v1.1Compatible | LeKiwiCompatible |
|---|---|---|---|---|
| GR00T N1.7A100 80 GB or H100 80 GB | Open guide90 min · intermediate | Open guide75 min · intermediate | Open guide90 min · advanced | Open guide100 min · advanced |
| GR00T N1.5A100 80 GB or H100 80 GB | Open guide75 min · intermediate | No guide yet | No guide yet | No guide yet |
| Pi0.5A100 80 GB or H100 80 GB | Open guide25 min · intermediate | Open guide18 min · intermediate | Open guide20 min · advanced | Open guide22 min · advanced |
| SmolVLARTX 4090 or any card with 24 GB | Open guide45 min · beginner | Open guide35 min · beginner | Open guide40 min · intermediate | Open guide50 min · advanced |
| ACTRTX 4090 or any card with 24 GB | Open guide25 min · intermediate | Open guide18 min · intermediate | Open guide20 min · advanced | Open guide22 min · advanced |
Which model should I pick
The arm rarely decides this, the task and your budget do. All five models read the same LeRobot v2.1 datasets, so a dataset you record once can be trained more than once.
Runs on a 24 GB card, so a full run costs a fraction of an A100 hour.
Pretrained on a multi robot corpus, so it copes with object positions your demonstrations never showed.
Flow matching gives continuous trajectories, at the highest latency of the five models.
No pretraining and no language conditioning, but by far the shortest control loop.
Kept available for exactly that case. New projects should not start here.
Still undecided
Read the two models against each other, or start at the model reference under /policies.
Start on the SO-100
The SO-100 is the reference arm of this platform, so these guides are the ones tested first and corrected most often. If you own an SO-101 or a LeKiwi, they still read correctly: the servos and the recording flow are the same.
How to train GR00T N1.7 on an SO-100
Fine-tune NVIDIA GR00T N1.7 on an SO-100 dataset: episode counts, LeRobot v2.1, the default hyperparameters, cost per run, and evaluation on the arm.
How to train GR00T N1.5 on an SO-100
Fine-tune NVIDIA GR00T N1.5 on an SO-100 dataset to reproduce a run from before N1.7: the conservative defaults of batch size 1, learning rate 1e-5 and 2000 steps, what an A100 hour costs, and how to evaluate against the original result.
Training Pi0.5 on an SO-100
Fine-tune Pi0.5 on an SO-100 dataset: prerequisites, the exact training defaults, what the run costs, and how 485 ms inference behaves on the real arm.
How to train SmolVLA on an SO-100
Record 30 episodes, start a SmolVLA run on a 24 GB GPU for a couple of dollars, and deploy the policy back to your SO-100. Full walkthrough with defaults.
How to train ACT on the SO-100
Train an ACT policy on an SO-100 arm: dataset rules, action chunking, chunk size and nActionSteps, the lerobot defaults, run cost and hardware evaluation.
Never recorded a dataset before? Start with the four tutorials under /learn, they end where these guides begin.
Not sure your data is good enough yet
A real SO-100 is online and free to drive in the browser. Try the motion you want to teach before you record fifty episodes of it.