GR-2

The Fourier GR-2 humanoid robot, developed by Shanghai-based Fourier (formerly Fourier Intelligence), marks a significant advancement in general-purpose humanoid robotics, launched in late September 2024 as the second-generation model in the GRx series. Building on the GR-1, the world's first mass-produced humanoid, GR-2 introduces substantial upgrades in hardware architecture, AI integration, and developer accessibility, emphasizing agility, precision, and real-world deployability. Architecturally, GR-2 stands at 175 cm tall, weighs approximately 63 kg, and boasts 53 degrees of freedom (DoF), enabling a wide range of human-like movements including torso rotation, grasping, walking, and floor-to-stand maneuvers. Its bionic design features a serial joint configuration (upgraded from GR-1's parallel setup) for simplified control, reduced maintenance, and seamless sim-to-real transfer. The robot employs seven distinct types of FSA 2.0 (Fourier Smart Actuators), each optimized for specific joint torque demands, delivering peak torques exceeding 380-436 N·m and equipped with dual-encoder systems (23/24-bit resolution) for doubled control accuracy in dynamic environments. The standout feature is the 12-DoF dexterous hands per arm, doubling previous dexterity, with six array-type tactile sensors per hand for real-time force sensing, object shape/material identification, and adaptive gripping. Each arm/hand supports a 3 kg payload. Mobility is enhanced with a maximum walking speed of 5 km/h, supported by stereo cameras for SLAM and pure vision perception, and a 6-axis IMU (3-axis acceleration + 3-axis gyroscope). Computing power includes an optional high-performance module with a 14-core Intel Core i7 CPU and 275 TOPS AI unit, running Ubuntu 20.04, with connectivity via Ethernet, USB3.0, and Type-C. Power is supplied by a detachable 950 Wh lithium-ion battery at 39.6 V nominal (46.2 V charge limit), offering ~2 hours runtime and quick-swap capability, constructed from aluminum alloy and engineering plastics for durability. On the AI front, GR-2 is 'Made for AI,' leveraging NVIDIA Isaac Lab (migrating from deprecated Isaac Gym), ROS, and Mujoco frameworks via the Fourier Toolkit SDK. This enables reinforcement learning (RL) for sim-to-real tasks, with NVIDIA TensorRT for inference optimization, CUDA for parallel processing, and cuDNN for PyTorch acceleration. Training demos show 89% success in floor-to-stand after 3,000 sim iterations in 15 hours, incorporating language models and predictive analytics for decision-making in unpredictable settings. Programming modes include VR teleoperation (Pico/VisionPro), lead-through teaching, and direct commands, capturing motion paths and tactile data for algorithm refinement. Regarding real-world deployments, as a prototype available for research, GR-2 is trusted by leading institutes and partners for applications in scientific research, healthcare (e.g., elderly assistance, rehab), manufacturing, services, and academia. No widespread commercial deployments yet, but it's positioned for high-interaction roles, with data collection advancing training for safe human-robot collaboration. Fourier emphasizes commercial viability through modular design, integrated cabling for compact packaging, and open interfaces, fostering embodied AI exploration across industries. (Word count: 612)