Train a policy
RL and VLA training, on cloud GPUs or your own machine. Every run tests its checkpoints in simulation as it goes.
What you can train
| Template | Policy | Needs | Runs on |
|---|---|---|---|
| MuJoCo PPOWalking and balance for a legged robot, reaching for an arm. The task comes from the robot's interface. | A robot with a declared interface. | CPU: your own machine, or Lucen cloud within the free allowance. | |
| Isaac Lab PPOThe Isaac Lab stack the Laika policies were trained with, for thousands of robots in parallel.By request | reflex | A robot with a declared interface. | An RT-core GPU: hosted for Lucen's robots, or your own machine with Isaac Lab. |
| SmolVLA LoRAFine-tunes SmolVLA, a 450 M parameter vision-language-action policy, on your own recordings. | A LeRobot dataset with a camera. | A GPU: your own, or a Lucen cloud tier. |
Where it runs
Lucen cloud CPU
MuJoCo training, robot checks and sim tests, within a free monthly allowance.
Free: $2 a month with an account.
Lucen cloud GPU
Metered by the second at the GPU's rate. Set a budget on each run, and the hub stops it there.
| GPU | Memory | Per hour |
|---|---|---|
| L4 | 24 GB | $1.024 |
| A10G | 24 GB | $1.408 |
| L40S | 48 GB | $2.496 |
| A100 | 80 GB | $3.20 |
| H100 | 80 GB | $5.055 |
Your own machine
Free, with no limit. Start a runner once; it takes your queued runs and trains them where it runs.
uv tool install "lucen-cli[rl]"
export PATH="$HOME/.local/bin:$PATH"
lucen auth login --url https://api.mousemouse.ai
lucen worker runWhat every run gets
Checkpoints as it trains
Each checkpoint is published to the policy's page, so you can test or deploy an earlier one.
The tutorialA sim test of each checkpoint
By default it tests the checkpoints that improved, then every one after the first pass. Each test scores Pass or Fail, with a video.
Reading a sim testStop rules
Stop at the first checkpoint that passes every test, or after a few tested checkpoints without improvement.
The train formA budget on every cloud run
The run stops when it reaches the budget you set. Its checkpoints so far are kept.
PricingThe Coach
Where this hub has an AI model, the Coach reads a finished run and proposes the next experiment. Each claim is cited.
The Training Coach
What runs cost
| Run | Cost | Took | Measured |
|---|---|---|---|
| Test a policy in simulationFive sim tests, two seeds each, on cloud CPU. | $0.0063 | 2 min 41 s | Measured on a hosted run, 3 October 2026. It fits in the free CPU allowance. |
| Train a MuJoCo policy150 PPO iterations on cloud CPU, with checkpoints saved as it trains. | $0.0364 | about 5 min | Measured on a hosted run, 3 October 2026. It fits in the free CPU allowance. |
| Train Laika to stand in Isaac Lab1,500 iterations, 4,096 robots in parallel, on one L40S GPU. | $1.64 | 30 min | Measured on our own run, 22 September 2026. Isaac Lab runs on our cloud only for Lucen's robots, under NVIDIA's licence. |
