RealSense: no frames or IMU with enable_imu=True under librealsense RSUSB (two contexts claim the same USB interfaces)

[alxbouchard]

real-sense-camera with enable_imu=True gets no frames and no IMU samples when librealsense is built with the RSUSB (libusb) backend.

Setup

• Jetson Orin Nano 8 GB, Ubuntu 24.04 / L4T R39.2.1 (kernel 6.8-tegra, no CONFIG_HID_SENSOR_*)
• dimOS installed with uv in a venv, commit faf1bed (same code path on current main)
• librealsense 2.57.7 built with -DFORCE_RSUSB_BACKEND=ON (the only way to reach the IMU without the hid-sensor drivers), realsense_native built with cargo against it
• D435i on USB 3, Zenoh transport

Command

dimos --viewer none run real-sense-camera --enable-imu true

What happens
realsense_native opens the camera twice: imu_open in its own context, then capture_thread in a second one. Under RSUSB each context claims the USB interfaces itself, so the second one finds them busy:

WARNING (handle-libusb.h:110) failed to claim usb interface, interface 0, is busy - retrying...
ERROR   (handle-libusb.h:125) failed to claim usb interface: 0, error: RS2_USB_STATUS_BUSY
ERROR   (uvc-sensor.cpp:428) acquire_power failed: failed to set power state

Then no frames within 1s / no IMU samples within 1s forever. Once the process segfaulted and the camera stopped enumerating (error -71) until replugged. With enable_imu=False video works. With the default V4L backend the kernel arbitrates the interfaces, which is probably why the D455 setup doesn't see this.

Proposed fix
One context and one pipeline for all streams (color, depth, IR, accel/gyro), started with a frame callback. Measured on the setup above: color/depth 14.97 Hz, IMU 399 Hz, clean stops 8/8. Stopping a callback pipeline over RSUSB takes ~2.3 s, so shutdown_timeout goes to 5 s.

Branch: https://github.com/alxbouchard/dimos/tree/alxbouchard/fix/realsense-single-context. I'll open it as a draft PR linked here. Happy to change the approach if you prefer something else.