PHYBOT C1

The PHYBOT C1 is a pioneering compact humanoid robot engineered by Beijing-based PHYBOT (Dongyi Technology), embodying the company's mission to 'Bring AI into the physical world' through co-designed hardware and software for embodied intelligence. Measuring 128 cm in height and weighing 28-30 kg, the C1 serves as a versatile platform for research, AI development, social companionship, elderly care, educational interactions, and light-duty assistance in indoor environments like homes, hotels, malls, and museums. Its architecture centers on the proprietary PhyArc series of quasi-direct-drive integrated cycloidal joint modules, marking one of the first uses of cycloidal actuators in small humanoids. These actuators feature 25:1 reduction ratios, dimensions from Ø47×68 mm to Ø102×54.8 mm, rated torques up to 60 Nm (peak 244 Nm for larger leg joints), powers up to 1080 W rated (5760 W peak), and maximum motor speeds of 15 rad/s. This design distributes loads across multiple rolling contacts, delivering high torque density, minimal backlash (<0.1°), exceptional durability, compact form factor, and smooth precision motion superior to traditional harmonic or planetary gears. The robot boasts 21-28 degrees of freedom (DoF): typically 3 hip (pitch/roll/yaw), 1 knee, 2 ankle (pitch/roll) per leg; 1 waist yaw; 3 shoulder (pitch/roll/yaw), 1 elbow per arm; with hands adding dexterity (up to 12 DoF total, 5 fingers each). Powered by a 72 V swappable battery pack offering 3-4+ hours runtime, it achieves walking speeds of 2-3 km/h, 5 kg single-arm payload, and 20 kg back load capacity. Compute is handled by an onboard Allspark-2 x86 platform with Intel Core i7-1165G7 processor, enabling 50 Hz policy inference and 500 Hz low-level PD control over wired CAN bus (<1 ms latency). Sensors include IMU, absolute encoders on joints, 1080p RGB + depth cameras, with external motion capture (e.g., FZMotion 21-camera system at 210 Hz) for advanced demos. Connectivity supports Bluetooth, Wi-Fi; OS is Linux-based ROS; LLM integration via external APIs; end-to-end latency 150-250 ms. AI capabilities shine in real-world deployments, exemplified by multi-stage reinforcement learning (RL) for whole-body badminton using PPO in IsaacGym (4096 parallel envs, 12.4 hours training on RTX 4090). The asymmetric actor-critic observes 87-D state + history (joint pos/vel/actions), predicts shuttlecock trajectories via EKF, and executes zero-shot sim-to-real transfers achieving 19.1 m/s peak returns, multi-hit rallies, and human-level play. Domain randomization (friction, pushes, inertia) ensures robustness. Other demos include autonomous hotel tasks, front handsprings, backflips, projector-face multimodal interactions (HD projection for movies/emotions, breathing lights), and dynamic warm-ups. Deployed at events like SWITCH 2025 Singapore (visited by dignitaries), research labs for locomotion/perception testing, and prototypes for companionship with skin-friendly composites/aluminum frame (IP20). As a prototype priced ~$22,000, C1 bridges affordable research hardware and commercial viability, fostering AGI-era productivity in AI labs, education, and service sectors.