AI Worker

The ROBOTIS AI Worker is a state-of-the-art semi-humanoid robot platform designed for enterprise-level industrial applications, emphasizing physical AI to automate complex tasks through human demonstration and learning. Standing at 162 cm tall with models like FFW-SG2 (mobile) and FFW-BG2 (fixed base), it features 25 or 19 degrees of freedom respectively, including dual 7-DOF arms (reach 641 mm to wrist), 1-DOF grippers (RH-P12-RN, 5 kg payload), a 2-DOF head, 1-DOF lift, and optional 6-DOF swerve drive mobile base capable of 1.5 m/s velocity. Powered by NVIDIA Jetson AGX Orin 32GB (Jetpack 6.2), it integrates ROS 2 Jazzy, ros2_control at 100Hz, and Dynamixel SDK for RS-485 communication (4 Mbps) with proprietary DYNAMIXEL actuators: YM series for shoulders/elbows/lift/wheels, PH for wrists, XH for head. Architecturally, the AI Worker employs a dual-system AI pipeline: a Vision-Language Model (VLM) for semantic understanding of visual and linguistic inputs, paired with a Diffusion Transformer (System 1) for generating action chunks from robot state, enabling end-to-end vision-language-action (VLA) control. It leverages NVIDIA Isaac GR00T N1.5 foundation model, fine-tuned on 10 hours of teleoperation data (800 episodes) using 8x B200 GPUs. Deployment uses Docker containers: AI Worker for ROS2 control, Physical AI Tools for inference orchestration via ZMQ/ROS2, and GR00T for GPU-accelerated prediction (40ms latency on RTX 5090, 10 FPS control). Inputs include RGBD cameras (Stereolabs ZED Mini head: 2208x1242, 0.1-9m depth, 102° FOV, 6DoF IMU; Intel RealSense D405 hands x2: 1280x720, 7-50cm depth, 87°x58° FOV), dual LiDARs, and joint/IMU feedback. Learning emphasizes imitation from teleop/VR demos and reinforcement for refinement, supporting tasks like wiring, welding, inspection, assembly, and logistics. Real-world deployments include CoRL 2025 and Humanoid Conference 2025 demos: 85% success in sorting coffee bottles amid crowds/lighting variations (failures: 50% misclassification, 30% grasp fails, 20% phantom actions). Swerve drive enhances omnidirectionality over mecanum wheels. Fully open-source (code, URDF, sims, datasets), it runs 4+ hours on 25V 80Ah (2040Wh) LiPo battery, weighs 90kg (SG2), uses aluminum/plastic frame, operates 0-40°C. Safety via joint limits, torque control. This bridges research to industry, addressing labor shortages via scalable Physical AI.