Ambidex

Ambidex, developed by Naver Labs in collaboration with Korea University of Technology & Education (KOREATECH), represents a breakthrough in dual-arm humanoid robotics, emphasizing safe human-robot coexistence through innovative cable-driven architecture. Initially unveiled in 2017 as lightweight robotic arms, Ambidex has evolved into a full upper-body humanoid platform standing 186 cm tall, incorporating a sensor-equipped head, torso, waist mechanism, and advanced grippers like the BLT Gripper. Each arm features 7 degrees of freedom (DOF), mirroring human arm kinematics, with actuators strategically placed at the shoulder and base to minimize distal mass—incredibly, each arm weighs just 2.6 kg, lighter than an average adult male arm. This cable-driven system uses wire-structured power transmission to achieve industrial-level precision, strength, and speed (up to 5 m/s tip velocity) while ensuring inherent safety through low inertia and compliant force control. The mechanism amplifies torque at joints, supporting payloads up to 3 kg per arm (upgradable to 5 kg in variants like LIMS3-AMBIDEX), enabling dynamic tasks such as high-fiving humans without injury. Architecturally, Ambidex employs a hybrid dynamic model for tendon-driven control, integrating physics-based simulation with data-driven learning to handle complex nonlinearities in cable tensions and joint interactions. The waist addition expands workspace, while the sensor head provides visual perception for object detection. Force/torque sensors embedded in joints and cables enable precise haptic feedback, crucial for bilateral teleoperation. Naver Labs' proprietary haptic devices (upgraded to 2.0) facilitate 1:1 human-scale control, transmitting position, velocity, and force bidirectionally with low latency (<5 ms cycles), allowing operators to 'feel' the robot's environment. This generates high-fidelity training data on 'physical intelligence'—subtle human motor skills like adjusting force when peeling vegetables or timing impacts in table tennis. AI integration centers on task learning via imitation and reinforcement learning (RL). Human demonstrations through haptics are replayed in simulators for policy optimization, enabling single-shot learning where Ambidex autonomously executes complex maneuvers like IKEA chair assembly, sweet potato peeling, ball catching, or pole balancing. Deep learning extracts behavioral patterns from interaction data, coupled with real-time motion planning, impedance control, and visual-tactile fusion. Cloud-based 'brainless' computing via 5G (1 ms latency with Qualcomm) offloads heavy processing, reducing onboard power (up to 40% savings) and enabling fleet learning from multiple units. Real-world deployments began with CES 2019 demos showcasing 5G remote control and have progressed to NAVER's 1784 headquarters—a robot-friendly building where Ambidex cleans service robots like Rookie, navigates dynamic environments, and collects petabytes of interaction data. Testbeds since 2017 in cafes, libraries, and logistics sites validate performance in unstructured settings, focusing on healthcare (rehab/nursing), domestic chores (cooking/laundry), manufacturing assembly, and entertainment. Ongoing evolution targets full autonomy through transfer learning and robustness in dynamic spaces, positioning Ambidex as a versatile platform for ambient intelligence ecosystems. With runtime up to 8 hours and connectivity via 5G/Wi-Fi, it bridges research and practical human augmentation. (728 words)