SmolVLA RGB-D Unfrozen B8 9280

SmolVLA policy fine-tuned for an SO-101 flower-pot manipulation task using three visual inputs.

Training configuration

Training steps: 9,280
Batch size: 8
Learning rate: 3e-6
Vision encoder frozen: No
Visual inputs: 3
Robot state dimension: 6
Action dimension: 6
Action chunk size: 50
Base VLM: HuggingFaceTB/SmolVLM2-500M-Video-Instruct

Dataset

The model was trained on a 200-episode RGB-D robot manipulation dataset.

Intended use

This model is intended for robotics research and demonstration. It requires the corresponding LeRobot/SmolVLA code, observation format, robot configuration, and hardware-specific calibration.

Limitations

The policy was trained for a specific robot, camera arrangement, workspace, and manipulation task. It should not be deployed on another physical robot without recalibration, safety limits, and supervised testing.