G2 Genie

The Agibot G2 Genie, launched on October 16, 2025, by AGIBOT Innovation (Shanghai) Technology Co., Ltd. (also known as Zhiyuan Robotics), represents a cutting-edge wheeled humanoid robot designed for industrial-grade, people-safe interactions. Standing at 175 cm tall and weighing 55 kg, it combines a humanoid upper body with an omnidirectional wheeled base for superior stability, mobility up to 7 km/h, and adaptability to over 95% of factory floor conditions. Its architecture features dual 7-degree-of-freedom (DoF) arms with full-arm high-precision torque sensing, a 3-DoF anthropomorphic waist for human-like bending, twisting, and swaying, and dexterous OmniHand Pro hands with 12 DoF and 10 fingers equipped with visuo-tactile fingertip sensors. This enables sub-millimeter precision assembly, force-controlled manipulation up to 5 kg per arm, and delicate tasks like RAM insertion or holding fragile objects such as raw eggs. Perception is achieved through a 360° spatial awareness system, including dual LiDARs for omnidirectional obstacle avoidance and hazard prediction, RGB-D cameras, ultrasonic sensors, and an 8-microphone array for multi-user voice conversations and eye-gaze tracking. Expressive facial animations and whole-body gestures enhance human-robot interaction, allowing the robot to explain actions to bystanders. At the core is an NVIDIA Jetson Thor T5000 compute platform delivering up to 2070 TFLOPS (FP4) or ~200 TOPS (INT8), enabling low-latency (<10 ms glass-to-action) on-device execution of large embodied AI models without cloud dependency. The AI stack leverages Agibot's proprietary WorkGPT/Genie multimodal mission-level models enhanced by Vision-Language Models (VLMs), integrated with the GO-1 generalist embodied foundation model featuring a '3-layer brain' architecture: VLM for perception, Latent Planner for task planning, and Action Expert for precise execution. The Genie Envisioner (GE-1) world model supports predictive simulation, policy learning, and neural rendering for rehearsing actions in virtual environments. Rapid deployment is facilitated by Genie RL reinforcement learning and LinkCraft zero-code platform, allowing non-experts to teach tasks via human demonstration in as little as one hour, with 99% success rates in precision assembly (e.g., completing components 3 seconds faster than humans in 15 seconds). Built with 100% automotive-grade components and IP42 ingress protection, it withstands extreme conditions (-15°C to 50°C), electrostatics, and includes dual hot-swappable batteries for 24/7 autonomous operation with self-charging. The wheeled design prioritizes industrial applications over bipedal locomotion, focusing on manipulation, guidance, and service tasks. Real-world deployments began shortly after launch, with commercial pilots in automotive manufacturing (e.g., seatbelt lock cylinder assembly and material handling alongside humans) and consumer electronics precision tasks. By late 2025, Agibot reported ongoing rollouts in these sectors, building on the company's first industrial deployment in July 2025 (earlier models) and mass production scaling to thousands of units. Agibot's AIDEA Giga Data Factory and AgiBot World dataset (over 1 million real-machine interactions) fuel continuous AI improvements, positioning G2 Genie as a benchmark for embodied AI in logistics, research, education, factories, retail, and public venues. Its force-compliant, interactive nature ensures safety in human-collaborative environments, marking a shift from rigid industrial arms to versatile, intelligent operators.