rewact_tools

RewACT Tools: Plugins and utilities for reward-based learning with LeRobot

Installation

pip install rewact_tools

Or install from source:

cd rewact_tools
pip install -e .

Features

This package provides:

• Plugins for robocandywrapper: Reward calculation, advantage computation, and control mode tracking
• Factory functions: Utilities for creating pre/post processors for RewACT and ACTvantage policies
• Dataset utilities: KeypointReward and LeRobotDatasetWithReward for managing reward labels

Plugins

• DenseRewardPlugin: Adds dense reward calculation with keypoint-based interpolation
• PiStar0_6CumulativeRewardPlugin: Computes cumulative rewards for pi*0.6 style training
• PiStar0_6AdvantagePlugin: Computes advantages for advantage-conditioned policies
• ControlModePlugin: Tracks control mode (human vs autonomous) in datasets

Usage

from rewact_tools import DenseRewardPlugin, KeypointReward, make_pre_post_processors

# Use plugins with robocandywrapper
# Use factory functions for policy processors
# Use KeypointReward for defining reward keypoints