Simulations and Visualisations
We keep trying new and interesting algorithms at Ati. Sometimes a one-off custom simulation is used to test them out and then we use our in-house unity based simulator to try more realistic scenarios. For example, look at this very simple and cool simulator that was written to try a probabilistic approach to collision avoidance. One can simply paint a path and obstacles using a paint brush to begin:
It did well enough to graduate to a more sophisticated simulator
Next up it should get tested on the physical vehicle itself!
Similarly, we were trying some imitation and reinforcement learning ideas. Here the first simulator itself was in unity since it was easy to generate random traffic on it.
Apart from simulations, we need to visualize our sensor data regularly. See here how we currently visualise our Lidar based object detection. The camera feed is on the side to correlate what we are interpreting. The data is being interpreted and translated into a top view. This is the view that the path planning module uses as an input. Later this will get fused with the camera based object detection for even superior results. Apart from the object itself, lidar data also gives us the crucial distance data.
The key idea is to have multiple competing techniques for the important objectives we have rather lock ourselves into just one approach. The best one’s win at the end. Or sometimes a cool hybrid approach emerges as a combination.