로봇과 AI의 변화를, 근거부터 쉽게.
What is a humanoid robot, and when is it worth deploying? Read public examples from Atlas, Figure, Gemini Robotics, Optimus and Unitree G1, then calculate throughput and cost per completed job—including failures and recovery. A practical starting point for robotics technology and business decisions.
Hand-eye calibration estimates the rigid pose between a robot frame and a camera from paired observations across several poses. This explainer covers AX=XB, motion diversity, validation and the practical limits introduced by timing, kinematic error, target quality and mechanical flex.
Null-space control lets a redundant robot preserve a primary task, such as holding its hand pose, while using remaining joint freedom for obstacle, joint-limit or singularity avoidance. We explain the Jacobian projector, practical priorities and why modeling error, latency and saturation can still disturb the primary task.
Task and motion planning links a robot’s discrete action sequence with grasps, placements and collision-free trajectories. Using a kitchen example, we explain how TAMP backtracks when geometry invalidates a plan—and why uncertainty, deformable objects and execution errors still require sensing and replanning.
Single-leg stance looks simpler than acrobatics, yet it requires a robot to track not only where its center of mass is but where it is moving. We examine DDC’s 89-of-90 simulation result and the important limits of its qualitative Unitree G1 demonstration.
Model predictive control forecasts a robot’s short-term motion, optimizes actions under explicit constraints, executes only the first input and then solves again. This explainer uses legged locomotion to show the principle—and the limits imposed by model error, computation deadlines and safety requirements.
A behavior tree breaks a complex robot mission into reusable conditions, actions and control-flow branches that report success, failure or running. Using a Nav2 recovery example, this explainer shows how the structure works—and why it should not be mistaken for a safety-rated system.
Skild AI says S1 can execute unseen, long-horizon robot tasks from one video demonstration without task-specific weight updates. This article explains its 66% versus 9% internal benchmark, the human-recovery condition behind that metric, and what still must be proven for reliable factory deployment.
Understand how motors, transmissions and sensors become humanoid joints. This guide explains brushed DC versus BLDC, strain-wave versus planetary gearing, and motor versus joint torque, then compares what public sources actually establish about Atlas, Figure and Optimus without mixing generations or counting methods.
A control barrier function can preserve a robot’s nominal command while minimally correcting inputs that would cross a defined safety boundary. This explainer covers safe sets, quadratic-program safety filters and the practical limits imposed by sensing delays, model error, infeasible constraints and independent safeguards.