ivelin/zk0-smolvla-fl

zk0 SmolVLA Federated Learning Model

This is a SmolVLA model trained using federated learning on SO-100 real-world robotics datasets.

Model Description

• Base Model: SmolVLA (lerobot/smolvla_base)
• Training Method: Federated Learning with Flower framework
• Strategy: FedProx (proximal_mu=0.01)
• Rounds: 30
• Clients: 4
• Local Epochs: 50 per round
• Batch Size: 64

Training Configuration

• Proximal Mu: 0.01
• Initial Learning Rate: 0.0005
• Evaluation Mode: Full
• Evaluation Frequency: Every round
• Checkpoint Interval: Every 10 rounds

Client Datasets

The model was trained on 4 different SO-100 robotics manipulation tasks:

1. lerobot/svla_so100_pickplace (replaced with shaunkirby/record-test in round 5)
2. lerobot/svla_so100_stacking
3. lerobot/svla_so100_sorting
4. lerobot/svla_so100_tool_manipulation

Evaluation Results

Final Server Evaluation (Round 30)

• Policy Loss: 0.544
• Action Dimensions: 6
• Successful Batches: 33/34
• Total Samples: 2112

Training Progression

• Initial Client Loss (Round 1): 2.534
• Final Client Loss (Round 30): 0.339
• Parameter Update Norm (Round 30): 1.437

Usage

from transformers import AutoModel, AutoConfig
import torch

# Load the model
model = AutoModel.from_pretrained("ivelin/zk0-smolvla-fl")
config = AutoConfig.from_pretrained("ivelin/zk0-smolvla-fl")

# The model is ready for inference on robotics manipulation tasks

Federated Learning Details

This model was trained using the Flower federated learning framework with:

• Privacy Preservation: No raw data leaves client environments
• FedProx Strategy: Addresses non-IID data with proximal regularization
• Server-Side Evaluation: Global model evaluated on unseen SO-100/SO-101 tasks
• Secure Communication: Parameter aggregation with integrity validation

Citation

If you use this model, please cite:

@misc{zk0-smolvla-fl,
  title={zk0 SmolVLA Federated Learning Model},
  author={Ivelin Ivanov},
  year={2025},
  url={https://huggingface.co/ivelin/zk0-smolvla-fl}
}

License

This model is released under the same license as the original SmolVLA model.