Make a Robot Arm Feel Weightless with NVIDIA Holoscan 5

Nvidia has published a walkthrough for building a gravity-compensation loop with Holoscan 5, its runtime that now pairs robotic control with sensor processing. The setup uses the Seeed Studio reBot Arm B601-DM, the Pinocchio inverse-dynamics library and the motorbridge motor library. The result is an arm that holds up its own weight while yielding to a person's hand, a starting point for teaching robots by demonstration.

The headline measurement: pinning the control partition to a dedicated CPU core cut tail latency at the p99.99 level by roughly 5x in C++ and 9.6x in Python. For anyone building robots that work near people, timing consistency is a safety property, not a nicety.

We like tutorials that come with measurements, and this one shows its work. The loop is three operators: one reads joint state every 2.5 milliseconds, or 400 Hz; one computes gravity torques with Pinocchio; one sends commands to the motors. Holoscan groups them into a single partition on one thread to avoid handoffs, and uses depth-one, drop-oldest queues so stale data never piles up. It's simple, and that's the point.

The latency table is the most useful thing here. Pinned C++ ran at a median of 7.1 microseconds and 17.2 at p99.99. Unpinned Python hit 1,373.5 microseconds at p99.99. That's an honest look at the cost of Python's interpreter and GIL in tight control loops, and Nvidia doesn't hide it. It also argues, sensibly, that Python remains practical for GPU perception work measured in milliseconds.

The safety design deserves credit. Commands that are expired, out of order or out of range get rejected, and a watchdog tries to disable the motors if no command batch completes within 100 ms. Holoscan also compiles an inspectable execution plan before motors are enabled, catching configuration errors before hardware moves. That's the habit we want across robotics: check before you power up.

Two cautions. This is vendor content, and the comparison is Nvidia's own runtime against its own configurations, so treat it as a strong tutorial rather than an independent benchmark. And Nvidia notes the core-isolation approach still lets interrupts, timers and some kernel threads run on reserved cores, so this isn't hard real-time.

Still, the trajectory is good. Accessible torque-controlled arms plus inspectable runtimes lower the bar for hobbyists and researchers to build compliant robots that are safer to touch. As AI policies start feeding goal positions into arms like this, a transparent, measurable control layer underneath is exactly where we want the rigor.

The walkthrough was tested on Jetson Orin Nano, Jetson Thor and x86 Ubuntu 24.04 systems. The reBot Arm uses quasi-direct-drive actuators with high torque bandwidth, unlike typical low-cost arms built on hobby servos suited only to position control. A second example in the tutorial adds kinematics and impedance control so an arm can hold a target pose while staying compliant to touch.