Robotics Briefing /
When should a robot give way?
A small walking robot gets pressed into the ground. A humanoid holds a bag while someone adds weight. In the first experiment, it helps the mechanism to give way. In the second, it helps the robot to push back. This weekend’s reading looks at how both teams got the response they wanted, and what happened when they put it to a physical test.
Your quick takeaways
- A six-legged robot with magnetic clutches recovered in 91% to 95% of trials after a probe pressed on its body, depending on how long the probe held it down.
- A humanoid that could adjust how hard it pressed its palms against a bin lifted the two-kilogram bin in nine of ten attempts.
- Stereolabs measured a 61% shorter wait for a large 3D camera frame to reach another program on the same GPU. Smaller frames saw less benefit.
Papers
A magnetic clutch lets a leg slip under load
Donato Romano and Gianluca Manduca · November 5 publication in npj Robotics · Physical prototype
On INSECTRON, a miniature six-legged walker, magnets connect each leg to the gears that drive it. The connection transmits the motor’s turning force during an ordinary step. When a leg meets too much resistance, the connection can slip. A second set of magnetic couplings between the drive modules helps the two parts of the body move back into alignment after a disturbance. The leg can also be replaced without tools.
The researchers pressed a rigid probe onto the robot at two locations and held it for a tenth of a second, half a second or a full second. Across 360 physical trials, the robot regained a stable configuration in 91%, 95% and 94% of trials at those respective durations. In some failures the legs struggled to stabilize. In others the two modules stayed apart after the probe lifted.
The distinction matters for a machine with little room for sensors: the clutch can release as soon as the load exceeds what the magnets will hold. The recovery figures apply to a press from above at the tested locations. A sideways impact or a leg caught under debris would impose a different load.
A humanoid changes its grip when a load changes
Mengjie Zhao and colleagues · November 3 preprint · Simulation and physical Unitree G1 trials
A robot can get both palms to the right spots on the sides of a bin and still let it slide. Ctrl-F adds another instruction to a movement: how much force to apply, and in which direction. A planner watches the object’s motion and the robot’s own body position, then updates the force command as the movement unfolds. It can respond to a change in weight without being told the weight itself.
On the physical humanoid, the authors report nine successful lifts in ten attempts with a two-kilogram bin. The comparison controller tracked the motion but its hands slipped off. In another test, someone added a kilogram to a one-kilogram bag hanging from the robot’s forearm. With its force planner active, the robot held its wrist at the target height. With the force command set to zero, the wrist dropped.
The hardware relied on printed markers to locate objects, and the controller learned a limited set of movements. Following a new movement would require retraining. The authors’ project page shows the physical trials and lists the code as forthcoming.
News
A shorter trip for a depth camera’s data
Stereolabs · November 5 technical report and November 7 ROS community announcement
A depth camera can produce a point cloud, a collection of 3D positions that describes nearby surfaces. Those points may already sit in the robot’s graphics processor when a mapping program needs them. Stereolabs’ version 5.6 camera software lets compatible programs on the same processor read that memory directly. The usual route involves copying the points out and then back again.
On a Jetson AGX Thor computer with a ZED 2i camera, Stereolabs timed the interval from publishing a frame to having it available on the receiving program’s GPU. It reports a 61% reduction for a full-resolution, 14.7 MB point cloud and 10% for a compact one. That saves time before mapping or obstacle detection can begin. The benchmark leaves open how much sooner a moving robot could react after those later steps. Programs that require ordinary CPU messages can still receive them.
Read this first
Watch the side-by-side bin lifts on the Ctrl-F project page. Follow the bin from the moment the palms meet it. The comparison robot gets its hands into place, but the bin slides as the hands rise. It’s a particularly clear illustration of why tracking the hands alone can miss the outcome of a grasp.
Read the paper