AgiBot (Zhiyuan Robotics) built the G2 to answer one question: what form factor wins in flat-floor factory environments? The answer is wheels. The G2 trades bipedal walking for an omnidirectional wheeled base, gaining operational stability, energy efficiency (8-hour battery vs 2-3 hours bipedal), and zero fall risk. While competitors demonstrate walking capability, AgiBot deployed G2 units at CATL and BYD, collected production data, and improved AI with every shift. The G2 is engineering optimized for what factories need rather than what looks impressive at trade shows.
Technical Specifications
- Type: wheeled humanoid, omnidirectional base + dual 7-DOF arms
- Height: adjustable 140-185 cm (motorized torso lift)
- Weight: 185 kg total system
- Arm DOF: 7 per arm = 14 total (shoulder 3, elbow 1, wrist 3)
- Hand DOF: 12 per hand = 24 total (5-finger dexterous)
- Force control: 0.5 N resolution at wrist
- Assembly precision: sub-millimeter (±0.5 mm) at TCP
- Mobile base speed: 1.5 m/s, 360-degree omnidirectional
- Battery: 8 hours continuous (vs 2-3 hours for bipedal platforms)
- Protection: IP42, -10°C to 50°C operating range
- Compute: NVIDIA Orin AGX + dedicated AI inference module
- Sensors: stereo RGB-D head camera, wrist cameras (both arms), 6-axis F/T sensors at wrist
- MES integration: MQTT, OPC-UA, REST
- Navigation: laser SLAM, pre-mapped route following, real-time obstacle avoidance
- Price: $95,000 at volume (reduced from $170,000 at 2024 launch)
Genie Sim 3.0 Training Platform
G2 manipulation AI is trained on Genie Sim 3.0, AgiBot’s simulation platform generating over 10 million synthetic scenarios in 2025. Built on NVIDIA Isaac Sim with photorealistic rendering and full environment randomization (lighting, object positions, surface textures, robot starting configurations). Sim-to-real transfer success rate above 85% for standard manipulation tasks. New task commissioning: hours rather than weeks. Data from CATL production deployments feeds back into Genie Sim to improve real-world performance continuously.
CATL Production Deployment, Task Breakdown
- Battery module quality inspection: wrist camera + AI detection of cell surface defects and tab weld quality
- Component transport: autonomous navigation on pre-mapped routes with variable payload
- Tray handling in formation/aging areas: precision pick-and-place from storage racks to testing fixtures
- Task success rate in production: above 95% for deployed task types
- Human intervention rate: below 5%, target below 1% by 2027
MES Integration Architecture
From the MES perspective, a G2 unit appears as a flexible work center with a task queue. Task assignment format: JSON payload specifying task type, target location ID, object description, and quality parameters. G2 returns completion status, quality check result, and exception flag. Integration timeline for a new factory deployment: 4-8 weeks including route mapping, MES connection, task commissioning, and operator training. The MQTT/OPC-UA interface is identical to conventional industrial robot integration, no specialized humanoid middleware required.
Commercial Status
- IDC ranking: #1 Chinese humanoid manufacturer by market share H1 2025
- Units delivered: 10,000 combined A2 + G2 by March 2026
- Total funding: $700M+
- IPO status: preparation ongoing as of mid-2026
AgiBot hardware comparison: AgiBot A2 vs G2. CATL deployment details: AgiBot at CATL.
