ROBNEWS
Twenty five years after the first documented use of robots in urban search and rescue, the field is shipping whole body agility and AI with publicly released model weights.
On September 11, 2001, the Center for Robot-Assisted Search and Rescue (CRASAR) deployed remotely operated machines to the World Trade Center site, the first documented use of robots in urban search and rescue. Personnel worked the pile through October 2, the day the last available robot failed. No survivors were found, but the machines located remains and mapped routes toward basements and stairwells.
Twenty-five years on, the field those machines helped create looks like a different species. IEEE Spectrum's Video Friday is a weekly curated roundup of robotics research, the kind of column a non-beat reader can use to map where the field actually is. This week's edition gathers five demos that look unrelated on the surface but share the same 2001 starting point.
A humanoid from ETH Zurich's Robotic Systems Lab swings across monkey bars, treating the bars as sparse 3D geometry it has to perceive and grip with its whole body rather than as a fixed ladder. A humanoid that can hang from a bar can, in principle, pull itself over rubble that defeated 2001-era machines. The roundup also features fresh legged-locomotion work, the part of the field most directly aimed at moving across the kind of broken ground that buried the World Trade Center site.
Unitree has fully open-sourced UnifoLM-WLA-1.0, an embodied foundation model: a single AI system that can drive many different robot bodies. The project site demos the same model coordinating desktop arms and a whole-body mobile platform, with cross-task and cross-end-effector generalization. Unitree says the release sets new state-of-the-art (SOTA) results across multiple benchmarks among open-source models worldwide. That is a self-claim, and a strong one; the SOTA framing belongs to Unitree, not to a third-party benchmark. What is independently true is that the weights are public, which means other labs can audit, fine-tune, and stress-test the same brain on their own hardware.
A third demo, from DRAGON Lab, makes the case for contact-rich work in the air. Their multilinked aerial robot uses centroid and joint motion for hybrid impedance-admittance control, a scheme that lets the platform both push and yield on contact. The application is surface sliding: a flying arm that can press against a wall and keep moving without bouncing off or stalling. DRAGON Lab says the work is accepted at IEEE IROS 2026, one of the field's main annual venues.
What ties the five demos together is a thread the CRASAR origin moment made visible. The 2001 robots were rugged, single-purpose, remotely operated machines. They had to survive dust, heat, and the structural mess of a real disaster. The 2026 versions are generalists: a humanoid that perceives sparse geometry, a foundation model that transfers across bodies, an aerial platform that handles contact without a human in the loop. The disaster context is no longer a separate category of robot. It is a stress test for the same general-purpose stack the rest of the field is shipping.
The venues will pile up fast. Humanoids Summit Seoul runs September 22 to 23, IROS 2026 in Pittsburgh runs September 27 to October 1, and CoRL 2026 in Austin runs November 9 to 12. By the time the field gathers at CoRL in November, the question will not be whether whole-body agility and open-weight embodied AI can demo in a lab. It will be whether any of it can keep working when the lab is a pile.