Custom Robot Trained in Isaac Sim/Lab

This project explores the cutting edge of sim-to-real robotics transfer using NVIDIA's Isaac Sim platform. By designing and training a custom robot entirely in simulation, the project demonstrates how virtual environments can accelerate robotics development and enable rapid experimentation with different designs and algorithms.

The work focuses on creating a photorealistic robot model, implementing domain randomization techniques, and training reinforcement learning policies that can potentially transfer to real hardware.

The software stack leverages NVIDIA Isaac Sim's USD-based simulation environment with Isaac Lab for reinforcement learning training. The robot model is defined using URDF and imported into the simulator with accurate physics properties.

Training uses Proximal Policy Optimization (PPO) implemented in PyTorch, with custom reward shaping for task-specific behaviors. The system includes ROS integration for potential hardware deployment and teleoperation capabilities. Domain randomization is applied to textures, lighting, and dynamics parameters to improve sim-to-real transfer.

Successfully trained policies for navigation and manipulation tasks, achieving convergence within reasonable training times. The simulation environment allows for 10x faster iteration compared to physical robot testing. The learned behaviors demonstrate robust performance across varied environmental conditions within the simulation.

Next steps include physical robot construction and validation of sim-to-real transfer performance.