Lifting with the whole arm, navigating from a room camera.

CALM: using the arm itself to lift a large box

Jun Hu and colleagues, Great Bay University and collaborators • September 24 • Weekend catch-up; preprint with physical tests

A box can exceed a gripper’s opening while still fitting between parts of a robot arm. CALM teaches a single UR7e arm to reposition a box, trap it between designated arm surfaces and lift it. The work explores handling bulky objects without a second arm or specialized tactile skin.

The team first trains separate skills in simulation, then combines their demonstrations into one controller. A learned model translates simulated joint histories into motor-current readings that resemble the physical robot’s readings. Those signals give the controller information about loading and contact.

In physical tests, CALM completed 22 of 30 trials. Success required holding the box at least twenty centimeters above the table for five seconds. A version without current alignment succeeded in none of ten trials; six of those trials triggered protective stops from excessive force.

The comparison supports the value of matching the simulated feedback to the hardware. It remains a small test with two cardboard boxes. Cameras tracked markers on the boxes, and the controller received their measured dimensions, so this does not establish handling of arbitrary warehouse packages.

The task controller learns in simulation, but the current-mapping model still uses physical recordings. Mobile-base motion is another unresolved source of changing motor loads.

• Paper ↗

ReVNM: planning a route from a camera across the room

Michikuni Eguchi, Kohei Honda and colleagues, Tsukuba, CyberAgent and Nagoya • September 24 • Weekend catch-up; preprint with physical tests

A fixed camera can see a large part of a room, but furniture can hide the space immediately ahead of a robot. ReVNM predicts what a depth camera on the robot would see, then uses that prediction with the remote view to choose waypoints.

The team trained the system in generated environments and transferred it to a different physical robot without further training. In an office layout with scattered obstacles, it completed 19 of 25 trials. In a second layout requiring passage through a wall opening, it completed all five trials.

The extra predicted view had mixed value. Removing it improved the scattered-obstacle result to 23 of 25, but reduced success at the wall opening to two of five. An inferred view can help when obstacles block useful information, while adding errors when the remote image already shows enough.

This approach could reduce the need for onboard visual processing and a prebuilt map. It still used onboard LiDAR for local obstacle avoidance and an AprilTag marker to locate the robot in the camera image.

Thirty physical trials across two layouts demonstrate an early transfer result. Reliable camera coverage and robot tracking remain requirements, and predicted depth can be inaccurate. The paper’s small wall test should not be read as a general 100% navigation rate.

• Paper ↗

IFR puts the world’s factory robot population at five million

International Federation of Robotics • September 24 • Catch-up: World Robotics 2026 report, covering 2025

The IFR reports that the number of industrial robots operating worldwide reached five million in 2025, up 9%. Factories installed more than 600,000 robots during the year, an 11% increase.

China accounted for 59% of new installations, with about 354,000 units. The United States installed almost 38,500, up 12%, moving ahead of Japan into second place. The report also shows uneven demand: installations fell in Japan and several major European markets.

These figures give the research stories a useful scale. A large installed base makes reliable integration, maintenance and upgrades consequential, even when an individual improvement looks modest.

The totals cover industrial robots, not a count of humanoids or AI-powered generalists. They also measure installations and operating stock, not uptime, productivity gains or jobs displaced. The September publication reports last year’s activity.

• IFR announcement and market charts ↗

Behavior Predictive Control: acting from stored demonstrations

Maximilian Adang, Timothy Chen and colleagues, Stanford • September 24 • Weekend catch-up; preprint with physical demonstrations

Training-free Behavior Cloning introduces Behavior Predictive Control, which retrieves and blends recorded demonstrations to choose actions. Keeping those examples available makes it possible to trace actions back to their source and revise behavior by editing the demonstration bank.

On an xArm6, it completed 34 of 40 trials moving a bag to the opposite plate, versus eight of forty for a fine-tuned π0.5 comparison. Remembering the bag’s starting side mattered. On the other two physical tasks, the comparison policy performed better: 39 versus 36 successes for coffee-bag placement and 40 versus 38 for a drawer task.

The system also ran onboard a drone at 75 updates per second on a Jetson Orin Nano. It completed ten of ten loop flights and eight of ten figure-eight flights using three and four demonstrations respectively.

The title’s training-free claim means no end-to-end policy training. Retrieval and correction components still require fitting. Missing demonstrations and visually similar situations can cause failures, and the paper provides no general safety or task-completion guarantee.

• Paper ↗