Realsense T265 and Indoor localization

I only had a conceptual understanding of SLAM (Simultaneous Localisation and mapping). I thought the recently launched Intel RealSense T265 tracking camera would be good way to get some hands on understanding of SLAM.

More precisely I wanted to see what are the outputs of the T265 cam and if it can help in indoor position of my favorite robot — the donkeycar.

Hardware setup of the robot:

For the purposes of this experiment, I used mounted the T265 on top of the donkeycar roll-cage and connected it to a Nvidia Jetson Nano devboard. (The standard donkey uses a raspberry-pi).

T265 mounted on a donkeycar

Side view of T265 test vehicle (note the Jetson Nano)

Visualizing the t265 pose outputs from a run in my living room

The below graph visualizes the T265 pose outputs (x and z co-ordinates) from the sample run.

There I have my question answered. It appears that pose output (SLAM outputs)means current co-ordinates w.r.t an origin. And if we plot all such pose outputs, it will be in the shape of the path taken by the vehicle.

Visualizing the T265 pose output from a sample run

The top left camera stream in the above video is from the T265’s left fish-eye lens. The bottom video stream is from an external camera. You can see that the pose data from T265 pretty accurately describes the trajectory taken by the donkeycar.

It appears that T265 pose outputs have very good resolution too. Note the patch taken around the piano bench are shown distinctly by the pose visualization.

Software setup

I started from the stock Jetson Nano sd card image. I had to make the i2c bus number related change mentioned by fei to get PCA9685 PWM outputs working.

I installed only the user level components of Realsense (librealsense) to get the tracking camera working (the kernel components are required only for the depth cameras) on Jetson Nano. I used Tawn Kramer’s donkeycar fork (t265 and path part and path_follower templates were very helpful). The examples from Marrku were very helpful too.

I had to make some additional installations to get the T265 python modules and get it working from the donkeycar stack. I will update this section with more detailed steps subsequently.