MenteeBot V3

MenteeBot V3, developed by Israel's Mentee Robotics, represents a significant advancement in full-size humanoid robots optimized for industrial warehouse automation and heavy lifting tasks. Standing at 175 cm tall and weighing approximately 65-70 kg, it boasts a payload capacity of 25 kg, enabling it to handle strenuous logistics operations that reduce human workload and injury risks. The robot's architecture is fully vertically integrated, featuring proprietary custom-built actuators that deliver up to three times the power density of competitors, paired with bespoke electronics, on-board sensors, and advanced fail-safe compute systems. At its core is a modular AI stack that eschews end-to-end vision-language-action (VLA) models in favor of specialized components for superior robustness and efficiency. High-level planning leverages a large language model (LLM) to interpret natural language instructions, decomposing complex tasks into executable code via a custom robotic API language. This includes task dependencies, success criteria, error detection, and recovery strategies using a 'try-catch' paradigm. The LLM operates intermittently (seconds per call), either onboard or cloud-based, enabling long-horizon compositional planning. Perception is powered by a sophisticated stack: NeRF/3D Gaussian Splatting (3DGS) for real-time 3D mapping and semantic localization; distilled OWLv2 for 2D object detection (4x faster); nano-SAM for lifting detections to 3D point clouds; stereo vision for dynamic obstacles; and topography projection for navigation, stair climbing, and grasping. The vision system provides 360° coverage via five head-mounted cameras, including fisheye side lenses and a rear camera, ensuring comprehensive environmental awareness. Low-level control employs reinforcement learning (RL) policies trained from scratch in simulators like Isaac Gym using novel Sim2Real techniques with heavy augmentations for domain randomization. These policies run at 40 Hz on dual NVIDIA Jetson Orin AGX processors, outputting motor torques via custom controllers. Inputs include joint states, IMUs, and task-specific data (e.g., paths, grasp poses), achieving near-100% success rates in locomotion and manipulation of rigid objects up to 25 kg. A standout feature is rapid task adaptation via imitation learning from a single demonstration video. Using NVIDIA's FoundationPose for trajectory reconstruction and automatic curriculum RL refinement, new skills are acquired in hours without engineering intervention, supporting on-site customization. Real-world deployment history includes a landmark November 2025 demonstration where two V3 units autonomously coordinated to move 32 boxes from eight variable-height piles to four racks in an unedited 18-minute run, showcasing stability, dexterity, and multi-robot collaboration without teleoperation. Earlier, a February 2025 teaser highlighted its hardware prowess at NVIDIA CES. As a prototype, no commercial deployments yet, but Mentee plans 2026 factory-floor leasing at under $10/hour, targeting logistics amid labor shortages. With hot-swappable batteries offering 3-4 hours runtime (extendable to 24/7), refined gait for industrial stability, and hands delivering 30 N pinch force per finger, MenteeBot V3 bridges simulation to reliable autonomy, positioning it for scalable industrial impact.