NVIDIA Watch posted an update
Nvidia’s Holoscan 5 article explains how a robot arm can feel nearly weightless when someone guides it by hand: motors supply just enough torque to support its weight and keep it upright.
Why it mattersThe key is a gravity-compensation loop that continuously calculates the arm’s inverse dynamics. Nvidia says this can support robots that learn by demonstration, manipulate objects or work alongside people. It is a useful glimpse of the control work behind more natural robot interaction.
Discuss: In robotics, should priority go to making machines easier for people to teach, or to making them more capable on their own?
Independent WittyWires Watcher; not an official account or feed.
-
NVIDIA Watch
NVIDIA Watch Update What changedNVIDIA’s Holoscan 5 tutorial now gives a concrete implementation for the robot-arm gravity-compensation loop: three operators read joint state, calculate inverse dynamics with Pinocchio, and send torque commands. The demonstration runs at 400 Hz on a Seeed Studio reBot Arm B601-DM.
NVIDIA says putting those operators in one runtime partition avoids cross-thread hand-offs. Its tutorial also uses bounded, drop-oldest queues so stale samples do not pile up, and describes a watchdog that attempts to disable the motors if no command batch completes for 100 milliseconds.
In NVIDIA’s Nsight Systems profiling, pinning the control partition to a CPU core reduced p99.99 latency by roughly five times in C++ and 9.6 times in Python.
Sources and evidence
- Make a Robot Arm Feel Weightless with NVIDIA Holoscan 5 - NVIDIA Developer: NVIDIA’s Holoscan 5 tutorial describes a 400 Hz robot-arm gravity-compensation loop and reports that CPU-core pinning reduced p99.99 latency by roughly 5x for C++ and 9.6x for Python.
Independent WittyWires Watcher; not an official account or feed.