---
language: en
tags:
- federated-learning
- flower
- smolvla
- robotics
- manipulation
- so-100
library_name: lerobot
license: apache-2.0
---

# SmolVLA Federated Learning Checkpoint

This model is a fine-tuned SmolVLA checkpoint trained using federated learning on SO-100 robotics datasets.

## Training Details

**Training Type**: Federated Learning (Flower Framework)
**Base Model**: lerobot/smolvla_base
**Timestamp**: 2025-11-01T08:33:27.949813
**Version**: 0.3.8

### Hyperparameters

- **Server Rounds**: 250
- **Local Epochs**: 20
- **Proximal μ**: 0.01
- **Initial Learning Rate**: 0.0001
- **Batch Size**: 64
- **Fraction Fit**: 0.3
- **Fraction Evaluate**: 1
- **Eval Frequency**: 1
- **Eval Batches**: 16
- **Checkpoint Interval**: 10
- **Dynamic Training Decay**: True
- **Scheduler Type**: cosine_warm_restarts
- **Adaptive LR Enabled**: True
- **Adaptive μ Enabled**: True

### Training convergence chart

[Wandb Run](https://wandb.ai/ivelin-eth/zk0/runs/zk0-sim-fl-run-2025-10-28_22-57-59/panel/9mt5s17ot?nw=nwuserivelineth)


![image](/static-proxy?url=https%3A%2F%2Fcdn-uploads.huggingface.co%2Fproduction%2Fuploads%2F63acff55073766d0850bb883%2F0sRD-7gIh-kMX2vVBHs9F.png)

### Training Datasets

- **shaunkirby/record-test**: Put the red LEGO in the bin
- **ethanCSL/direction_test**: turn to the right side - VALIDATED CLEAN
- **gimarchetti/so101-winnie-us5**: rub the plush toy with the bottle - VALIDATED CLEAN
- **olingoudey/so101_stationary_mug445**: Put the stuffed animal in the mug. - VALIDATED CLEAN
- **sebastiandavidlee/chopsticksTable2Organizer**: Grasp each utensil, classify it, and place it in its designated organizer slot. - VALIDATED CLEAN
- **giovipeg/record-test**: Pick and place pink knob. - VALIDATED CLEAN
- **VoicAndrei/so100_cubes_three_cameras**: Pick up the green cubes from the white plate and place them in the black recipient. - VALIDATED CLEAN
- **qm30631122/so101_grab_pen_20sec_50ep_SCB_102725_2**: Grab pen and put in to cylinder. - VALIDATED CLEAN
- **wsi-dev/tictactoe**: Place the X in the top right box. - VALIDATED CLEAN
- **KeisukeSato1024/so101_pp_green_box**: Pick and place the green box. - VALIDATED CLEAN
- **stanl1y/record_60_simple**: Grab the red pen and place it into the bin. - VALIDATED CLEAN
- **RickRain/Imperio**: Grab the block and drop it on the black square. - VALIDATED CLEAN

### Evaluation Datasets

- **Hupy440/Two_Cubes_and_Two_Buckets_v2**: Pick up a cube. Is the cube red, put it in the white bucket. Is the cube white, put it in the red bucket. - VALIDATED CLEAN
- **dll-hackathon-102025/oct_19_440pm**: Put orange slice in glass. - VALIDATED CLEAN
- **shuohsuan/grasp1**: grasp the red cube and put it into the bin. - VALIDATED CLEAN
- **yinxinyuchen/p_p_2**: Grab the green cube and put the cube in the green box. - VALIDATED CLEAN
- **dll-hackathon-102025/oct_25_umbrella_435pm**: put umbrella in glass. - VALIDATED CLEAN

### Final Evaluation Metrics

- **Composite Eval Loss**: 0.16521226037293674
- **Aggregated Client Metrics**: {'avg_client_loss': 0.2547603279352188, 'std_client_loss': 0.11142692836996185, 'avg_client_proximal_loss': 0.01709086100260417, 'avg_client_grad_norm': 1.4328387384613357, 'num_clients': 3, 'param_update_norm': 1.9328867197036743}
- **Individual Client Metrics**: 3 clients
### Per-Dataset Results

- **Hupy440/Two_Cubes_and_Two_Buckets_v2**: Loss 0.1347
- **dll-hackathon-102025/oct_19_440pm**: Loss 0.1867
- **shuohsuan/grasp1**: Loss 0.1479
- **yinxinyuchen/p_p_2**: Loss 0.1741
- **dll-hackathon-102025/oct_25_umbrella_435pm**: Loss 0.1828

### Training Insights

- **Convergence Trend**: Policy loss: 0.1647 → 0.1652
- **Avg Client Loss Trend**: Started at 0.0000, ended at 0.2093
- **Param Update Norm Trend**: Average 1.450677
- **Client Participation Rate**: Average 3.0 clients per round
- **Anomalies**: Client dropouts in rounds: [0]

## Usage

This model can be used for robotics manipulation tasks. Load with:

```python
from lerobot.policies.smolvla import SmolVLAPolicy

policy = SmolVLAPolicy.from_pretrained("ivelin/zk0-smolvla-fl")
```

### Model Format
- Weights saved in secure safetensors format (model.safetensors).
- Load with: `from safetensors.torch import load_file; state_dict = load_file("model.safetensors")`
- Avoid legacy pytorch_model.bin for security reasons.

## Limitations

- Trained on SO-100 datasets only

## Citation

If you use this model, please cite:

```
@misc{zk0-smolvla-fl-2025,
  title={SmolVLA Federated Learning on SO-100},
  author={Kilo Code, Grok AI, ivelin.eth, and contributors},
  year={2025},
  url={https://github.com/ivelin/zk0}
}
```