ACT Policy for Gearbox Assembly
This model is an Action Chunking with Transformers (ACT) policy trained for a gearbox assembly task in Isaac Lab.
Model Details
- Architecture: ACT (Action Chunking with Transformers)
- Task: Gearbox Assembly
- Framework: Isaac Lab / PyTorch
- Dataset:
sim_gearbox_assembly_demos_filtered
Training Data Usage
The model was trained using all available episodes from the provided dataset.
Important Data Characteristics
- Process Simplification: The pin insertion step was omitted in all training episodes. The task focuses on the steps following pin insertion (e.g., gear mounting, cover installation).
- Data Quality Note: The training set deliberately includes one known failed episode where a pin fell over during the process, causing the subsequent cover installation to fail. This episode was not filtered out and was used in training.
Hyperparameters
The model was trained with the following configuration:
| Parameter | Value |
|---|
| Policy Class | ACT |
| KL Weight | 10 |
| Chunk Size | 100 |
| Hidden Dim | 512 |
| Feedforward Dim | 3200 |
| Batch Size | 32 |
| Learning Rate | 1e-5 |
| Epochs | 10000 |
| Seed | 0 |
| Save Interval | 200 |
| Early Stopping | 3000 |
Usage
This model is designed to be loaded with the ACT policy wrapper in the Galaxea Lab framework for the GalaxeaLab-GearboxAssembly environment.
