Industrial

Physical AI & Humanoid Robotics in a Real-World Lab

How can highly complex humanoid robots with many degrees of freedom be prepared for safe and efficient use in dynamic production environments? In the ARENA2036 research hall, the use of Unitree humanoids is being evaluated in collaboration with member Waveye. The goal is to develop and validate motion strategies and physical intelligence (Physical AI) in a virtual environment before the robots make their first physical contact on the shop floor.

ROS/ROS 2 Control Architecture & Policy Training

The application is based on a precise digital representation of the robot's geometry and dynamics using URDF description files. It connects to the control architecture via an interface network in ROS 2.

This control channel continuously exchanges motion commands, joint torques, and sensor data (IMUs and force-torque sensors). This enables the training of reinforcement learning models while accounting for real-world physical parameters such as friction, inertia, and shifts in the center of mass, without any risk of damage to the actual mechanical system.

Physics Simulation in NVIDIA Isaac Sim & Omniverse

The simulation and visual interaction are performed in NVIDIA Isaac Sim within the Omniverse platform. The digital twin processes the ROS control signals in real time and maps the dynamic behavior of the humanoid robot within the physical context of the research hall.

Thanks to this Sim2Real pipeline, balance algorithms, obstacle avoidance, and collaborative tasks are validated in a photorealistic 3D environment. This drastically accelerates development cycles and ensures safe human-robot interaction in manufacturing.