TASER — Two-Armed Self-Equilibrating Robot

Overview

An Isaac Sim simulation of a self-balancing wheeled robot with arms, combining navigation, reinforcement learning-trained locomotion, and inverse kinematics-based manipulation.

Navigation

TASER has two navigation stack implementations:

	Obstacle Representation	Path Planner	Tracking Controller
Left	Occupancy Grid	Distance Transform (from the Robotics Toolbox)	Pure Pursuit
Right	Polygonal Obstacles	RRT* with Dubins Paths	Model Predictive Control (MPC)

RRT* Path Planner

The RRT* path planner is implemented in C++ and generates Dubins paths using geometric primitives.

Locomotion

To achieve self-balancing locomotion, TASER is trained using reinforcement learning with Isaac Lab and Proximal Policy Optimization (PPO). Two tasks are defined:

1. Balance (top): The joint velocity action range is kept small to ensure stable balancing while the robot is at an idle state.
2. Track Velocity (bottom): The joint velocity action range is increased to allow the robot to track a desired velocity of the shape (lin_x, ang_z) while maintaining balance.

For both tasks, the robot arms' joint positions and velocities are randomized at the start of each episode to accommodate different center of mass configurations.

Manipulation

Each of the robot's arms consists of 3 revolute joints around the axes: y, z, y.

The pick manipulation task is implemented with simple inverse kinematics. First, the arms reach toward the target by computing desired joint positions and tracking them with a PD controller, then the picking motion is executed by commanding joint velocities along the y- and z-axes.

Usage

• Create a .env file in the docker/ directory based on the provided .env.template.

• Run the project's docker container either in a VSCode Dev Container or using:

  docker compose -f docker/compose.yaml run --build taser bash

• Launch the Isaac Sim simulation:

  sim_isaac
  
  # or
  omni_python src/taser/isaacsim/sim.py