This repository contains code and datasets used to study how different action space designs affect robotic manipulation policies.
We provide:
- teleoperated manipulation datasets collected on AgileX robots
- training and evaluation code for policy learning
- tools for running experiments across different action space parameterizations
- a cross embodiment dataset for transfer experiments
The goal is to provide a simple benchmark for analyzing how action representations influence policy performance in robot manipulation.
conda create -n das python=3.10
pip install -e .
bash train.sh
bash server.sh
On AgileX-PiPER:
python agilex/client_air_eef6d_align_init.py