Adam Lite

The Adam Lite humanoid robot, developed by PNDbotics, a Beijing-based company specializing in full-stack humanoid robotics, represents a significant advancement in affordable, modular humanoid platforms. Launched as a prototype around 2023-2024, Adam Lite stands at 160 cm tall and weighs 60 kg, featuring 25 degrees of freedom (DoF) distributed as 6 DoF per leg (3 hip, 1 knee, 2 ankle), 5 DoF per arm (3 shoulder, 2 elbow), and 3 DoF in the waist. Its core innovation lies in the use of 25 proprietary quasi-direct drive (QDD) force-controlled PND actuators, which provide high back-drivability, sensitivity, and up to 340 Nm torque in leg joints, enabling exceptional mobility and stability for dynamic tasks. The biomimetic design emulates human skeletal structure, with a serial hip joint configuration approximating ball-in-socket motion for natural range. Architecturally, Adam Lite employs a fully modular, reusable framework, integrating PND Smart Actuators (PSA series) that combine motors, reducers, encoders, drivers, and communication units. The PND Robot Control Unit (RCU) serves as the central hub, managing actuators, Battery Management System (BMS), power distribution, and a 16-port Gigabit Ethernet switch for real-time sensory data and control via the PND Network. Onboard computing is powered by a 12th-generation Intel i7 NUC processor, supporting low-latency execution of complex policies. Sensors include an Inertial Measurement Unit (IMU) for linear/angular velocities, gravity orientation, and base frame data, alongside actuator encoders for joint positions/velocities. Vision and dexterous hands are optional add-ons; the Lite variant uses wrist-mounted rubber balls for basic manipulation. AI integration is a cornerstone, leveraging PNDbotics' proprietary Deep Reinforcement Learning (DRL) stack with sim-to-real transfer. The system utilizes NVIDIA Isaac Gym for parallel training across 4096 environments, employing Proximal Policy Optimization (PPO) and Adversarial Motion Priors (AMP) for imitation learning from human mocap datasets (SFU, CMU, custom). Observations encompass 91 dimensions (IMU, joints, commands, phases), actions are 25D PD targets. Domain randomization (mass, CoM shifts, motor strength, forces, noise) bridges simulation gaps, yielding human-like gaits: straight-knee walking, heel-toe running, arm swinging, terrain adaptation, and disturbance recovery. Rewards emphasize commands, periodicity (von Mises for phases), symmetry, and regularization for safety/smoothness. Real-world deployments highlight its prowess. Early 2025 demos showcased blind locomotion over uneven terrain and RL policies for challenging environments. At WAIC 2025, Adam Lite evolved variants like Adam and Adam-U performed synced dances via motion capture. It debuted at VOYAGEX Music Festival (July 2025), playing keytar live. Competed in World Humanoid Robot Games (Aug 2025), completing 100m obstacle race—the only full-size humanoid to do so. Further showcases at IROS, CORL, WRC, and office assistant demos underscore versatility. Partnerships with Noitom Robotics and Tianyong Engineering accelerate commercialization. Despite prototype status, priced around $120,000, its open-source SDK, stability, and cost-efficiency position it for research, industry, and AGI integration, setting standards in embodied intelligence.665054