92 lines
3.2 KiB
Plaintext
92 lines
3.2 KiB
Plaintext
# Use SmolVLA
|
||
|
||
SmolVLA is designed to be easy to use and integrate—whether you're finetuning on your own data or plugging it into an existing robotics stack.
|
||
|
||
<p align="center">
|
||
<img src="https://cdn-uploads.huggingface.co/production/uploads/640e21ef3c82bd463ee5a76d/aooU0a3DMtYmy_1IWMaIM.png" alt="SmolVLA architecture." width="500"/>
|
||
<br/>
|
||
<em>Figure 2. SmolVLA takes as input a sequence of RGB images from multiple cameras, the robot’s current sensorimotor state, and a natural language instruction. The VLM encodes these into contextual features, which condition the action expert to generate a continuous sequence of actions.</em>
|
||
</p>
|
||
|
||
### Install
|
||
|
||
First, install the required dependencies:
|
||
|
||
```python
|
||
git clone https://github.com/huggingface/lerobot.git
|
||
cd lerobot
|
||
pip install -e ".[smolvla]"
|
||
```
|
||
|
||
### Finetune the pretrained model
|
||
Use [`smolvla_base`](https://hf.co/lerobot/smolvla_base), our pretrained 450M model, with the lerobot training framework:
|
||
|
||
```python
|
||
python lerobot/scripts/train.py \
|
||
--policy.path=lerobot/smolvla_base \
|
||
--dataset.repo_id=lerobot/svla_so100_stacking \
|
||
--batch_size=64 \
|
||
--steps=200000
|
||
```
|
||
|
||
<p align="center">
|
||
<img src="https://cdn-uploads.huggingface.co/production/uploads/640e21ef3c82bd463ee5a76d/S-3vvVCulChREwHDkquoc.gif" alt="Comparison of SmolVLA across task variations." width="500"/>
|
||
<br/>
|
||
<em>Figure 1: Comparison of SmolVLA across task variations. From left to right: (1) asynchronous pick-place cube counting, (2) synchronous pick-place cube counting, (3) pick-place cube counting under perturbations, and (4) generalization on pick-and-place of the lego block with real-world SO101.</em>
|
||
</p>
|
||
|
||
|
||
### Train from scratch
|
||
|
||
If you'd like to build from the architecture (pretrained VLM + action expert) rather than a pretrained checkpoint:
|
||
|
||
```python
|
||
python lerobot/scripts/train.py \
|
||
--policy.type=smolvla \
|
||
--dataset.repo_id=lerobot/svla_so100_stacking \
|
||
--batch_size=64 \
|
||
--steps=200000
|
||
```
|
||
You can also load `SmolVLAPolicy` directly:
|
||
|
||
```python
|
||
from lerobot.common.policies.smolvla.modeling_smolvla import SmolVLAPolicy
|
||
policy = SmolVLAPolicy.from_pretrained("lerobot/smolvla_base")
|
||
```
|
||
|
||
## Evaluate the pretrained policy and run it in real-time
|
||
|
||
If you want to record the evaluation process and safe the videos on the hub, login to your HF account by running:
|
||
|
||
```python
|
||
huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
|
||
```
|
||
|
||
Store your Hugging Face repository name in a variable to run these commands:
|
||
|
||
```python
|
||
HF_USER=$(huggingface-cli whoami | head -n 1)
|
||
echo $HF_USER
|
||
```
|
||
Now, indicate the path to the policy, which is `lerobot/smolvla_base` in this case, and run:
|
||
|
||
```python
|
||
|
||
python lerobot/scripts/control_robot.py \
|
||
--robot.type=so100 \
|
||
--control.type=record \
|
||
--control.fps=30 \
|
||
--control.single_task="Grasp a lego block and put it in the bin." \
|
||
--control.repo_id=${HF_USER}/eval_svla_base_test \
|
||
--control.tags='["tutorial"]' \
|
||
--control.warmup_time_s=5 \
|
||
--control.episode_time_s=30 \
|
||
--control.reset_time_s=30 \
|
||
--control.num_episodes=10 \
|
||
--control.push_to_hub=true \
|
||
--control.policy.path=lerobot/smolvla_base
|
||
|
||
```
|
||
|
||
Depending on your evaluation setup, you can configure the duration and the number of episodes to record for your evaluation suite.
|