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.

Downloads last month
15
Video Preview
loading