Robotics & Physical AI

Overview

Robotics & Physical AI at TNG covers foundation-model fine-tuning, reinforcement learning, simulation environments, and edge inference for real-world robotics. We bring autonomous robots into production — from hardware selection to safety measures.

Robot Foundation Models

We fine-tune state-of-the-art robot foundation models such as pi0.5 and similar Vision-Language-Action (VLA) models. Using Virtual-Reality teleoperation setups, we collect demonstration data that enables robots to learn complex manipulation tasks like folding laundry, object sorting, and assembly.

Our approach combines:

• Imitation learning from VR-teleoperation demonstrations
• Reinforcement learning with reward shaping and policy optimization
• Sim-to-real transfer via simulation environments and digital twins
• On-edge inference for real-time control on resource-constrained hardware

Research Projects

Humanoid Robots

We work with the Unitree G1 humanoid robot for teleoperation and reinforcement learning experiments. Our research focuses on whole-body control, manipulation, and human-robot interaction.

Robot Dogs

We use the Unitree Go2 robot dog for navigation, perception, and autonomous exploration tasks. Projects include SLAM (Simultaneous Localization and Mapping), obstacle avoidance, and path planning.

ROS 2 Development

We co-develop the Robot Operating System (ROS) version 2 and build ROS/ROS2 applications for a variety of robotic platforms.

Showcase Projects

• OpenArm — An open-source robotic arm capable of folding laundry, trained on 100 VR-teleoperation demos using a pi0.5 Robot Foundation Model
• Robo-Kart — A humanoid robot driving a go-kart, demonstrated at the Deutsches Museum München
• Dance with robots — Teaching a robot dance moves using reinforcement learning

Hardware & Safety

• Choice, validation, and procurement of robot hardware
• Physical safety measures for human-robot interaction
• Sensor networks and real-time kinematics
• Edge AI processing on compact computers (Raspberry Pi, NVIDIA Jetson Nano)
• IoT and automation integration