IEEE Spectrum’s weekly robotics video selection highlights work on humanoid balance, robot flexibility, and their ability to adapt to their environment. Some sequences are presented humorously, others with details of the technical architecture that makes them possible.
The editors first highlight a video from Agility Robotics, whose sequence of dragging a sofa is deemed impressive. They also note ReST-RL, a hierarchical reinforcement learning architecture, a method where the robot learns through trial and error. This system explicitly separates locomotion from stabilizing the carried load. Deployed on the humanoid Unitree G1, this modular approach shows, according to the presentation, a “zero-shot” generalization from simulator to reality, meaning it works without additional training on the physical robot, across different objects and external disturbances.
Balance, Loads, and Adaptation
The company Figure also posted an update on scaling up its computations. It states: “Solving the question of a robot in every home is not a data and compute problem, it’s a safety and cost problem.”
Humanoids are often impressive on screen, but most of their movements are, according to the source, carefully choreographed. Researchers at the Safe AI Lab of Carnegie Mellon University are working on the opposite: adaptability. Their system, called APEX, allows a humanoid to overcome obstacles through adaptive whole-body movements.
Researchers from North Carolina State University have also designed soft robots shaped like water droplets, which leap upward or forward when exposed to infrared light and continue jumping as long as the light is present. The body is made of a liquid crystal elastomer ribbon, to which a thin V-shaped aluminum tube is attached. Under light, the ribbon’s surface contracts and the ribbon twists, storing energy until a critical point, where the V snaps against the ground and propels the robot into the air.
Sensitive Skin and Harvesting in the Heat
The University of Tokyo presents a three-layer skin designed to cover the joints of its musculoskeletal humanoid Musashi-W. It features 44 elements sensitive to pressure and stretching, to mimic the role of skin in proprioception, the body’s perception of its own position.
In Turpan, China, nicknamed the “Fire City,” ground temperatures in summer can exceed 50°C. During the grape harvest, the quadruped robot Lynx M20S from DEEP Robotics carries heavy baskets, whereas farmers traditionally do this by hand under intense heat.
Finally, the OH! GYM! project by ROBOTIS brought together a first cohort of undergraduate and master’s students, who developed their own humanoid behaviors on the open-source Sapiens K1 platform. In one month, they followed the complete process, from creating movements in simulation to transferring them to a physical robot.
The selection’s calendar mentions the DARPA Triage Challenge Finals, scheduled for November, and then the CoRL 2026 conference, from November 9 to 12, 2026, in Austin.
