How do you teach your JetBot new tricks? In this post, we highlight NVIDIA Isaac Sim simulation and training capabilities by walking you through how to train the JetBot in Isaac Sim with reinforcement learning (RL) and test this trained RL model on NVIDIA Jetson Nano with the real JetBot.
The goal is to train a deep neural network agent in Isaac Sim and transfer it to the real JetBot to follow a road. Training this network on the real JetBot would require frequent human attention. Isaac Sim can simulate the mechanics of the JetBot and camera sensor and automate setting and resetting the JetBot. The simulation also gives you access to ground truth data and the ability to randomize the environment the agent learns on, which helps make the network robust enough to drive the real JetBot.
Training JetBot in Isaac Sim
First, download Isaac Sim. Running Isaac Sim requires the following resources:
- Ubuntu 18.04
- NVIDIA GPU (RTX 2070 or higher)
- NVIDIA GPU Driver (minimum version 450.57)
Next, apply this Isaac Sim patch.
For more information about how to train the RL JetBot sample in Isaac Sim, see Reinforcement Training Samples.
Isaac Sim can simulate the JetBot driving around and randomize the environment, lighting, backgrounds, and object poses to increase the robustness of the agent. Figure 3 shows what this looks like during training:
After being trained, JetBot can autonomously drive around the road in Isaac Sim.
Here’s how you can test this trained RL model on the real JetBot.
Figure 6 shows what the real JetBot is seeing and thinking.
Running the trained RL model on the real JetBot
To build a JetBot, you need the following hardware components:
- A brain: Jetson Nano Developer Kit (either 2GB or 4GB)
- A body: Waveshare JetBot Kit
- Memory: MicroSD card
- 18650 rechargeable batteries for the JetBot.
- A Wi-Fi dongle if you’re using the 2GB Jetson Nano. The 4GB Jetson Nano doesn’t need this since it has a built in Wi-Fi chip.
For more information about supported components, see Networking.
Assemble the JetBot according to the instructions. On the Waveshare Jetbot, removing the front fourth wheel may help it get stuck less. Also, the 2GB Jetson Nano may not come with a fan connector.
Flash your JetBot with the following instructions:
Put the microSD card in the Jetson Nano board. Plug in a keyboard, mouse, and HDMI cable to the board with the 12.6V adapter. Boot up and follow the onscreen instructions to set up the JetBot user.
Update the package list:
$ sudo apt-get update
If you are using the 2GB Jetson Nano, you also need to run the following command:
$ sudo apt-get dist-upgrade
After setting up the physical JetBot, clone the following JetBot fork:
$ git clone https://github.com/hailoclu/jetbot.git
Launch Docker with all the steps from the NVIDIA-AI-IOT/jetbot GitHub repo, then run the following commands:
These must be run on the JetBot directly or through SSH, not from the Jupyter terminal window. If you see
docker: invalid reference format, set your environment variables again by calling
Running the following two commands from the Jupyter terminal window also allows you to connect to the JetBot using SSH:
apt install openssh-server ssh email@example.com
After Docker is launched with ./enable.sh $HOME, you can connect to the JetBot from your computer through a Jupyter notebook by navigating to the JetBot IP address on your browser, for example, http://192.168.0.185:8888. If the setup succeeded without error, the IP address of the JetBot should be displayed on the LED on the back of the robot.
Unplug the keyboard, mouse, and HDMI to set your JetBot free.
Install stable-baselines by pressing the plus (+) key in the Jupyter notebook to launch a terminal window and run the following two commands:
apt install python3-scipy python3-pandas python3-matplotlib python3 -m pip install stable_baselines3==0.8.0
Upload your trained RL model from the Isaac Sim best_model.zip file with the up-arrow button.
You can also download the trained model. Launch the jetbot/notebooks/isaacsim_RL/isaacsim_deploying.ipynb notebook. Select each Jupyter cell and press Ctrl+Enter to execute it. The second cell for
PPO.load(MODEL_PATH) might take a few minutes. [*] means the kernel is busy executing. When it’s done, it changes to a number.
Running the camera code should turn on the JetBot camera.
Executing this block of code lets the trained network run inference on the camera and issue driving commands based on what it’s seeing.
To interrupt the while loop, choose Stop. To stop the robot, run
Sim to real work
We specifically tailored the training environment to create an agent that can successfully transfer what it learned in simulation to the real JetBot. When we initially created the camera, we used default values for the FOV and simply angled it down at the road. This initial setup did not resemble the real camera image (Figure 12). We adjusted the FOV and orientation of the simulated camera (Figure 13) and added uniform random noise to the output during training. This was done to make the simulated camera view as much like the real camera view as possible.
We originally trained using the full RGB output from the simulated camera. However, we found that it took several hundred thousand updates to the network for it to start driving consistently. To shorten this, convert all images from RGB to grayscale. You should see the network start to display consistent turning behavior after about 100k updates or so. If you see the reward plateauing after a few hundred thousand updates, you can reduce the learning rate to help the network continue learning.
We also wanted to create an agent that didn’t require a specific setup to function. Differences in lighting, colors, shadows, and so on means that the domain your network encounters after being transferred to the real JetBot is quite large. You can’t simulate every possibility, so instead you teach the network to ignore variation in these things. You do this by periodically randomizing the track, lighting, and so on. You also spawn random meshes, known as distractors, to cast hard shadows on the track and help teach the network what to ignore. This process is known as domain randomization and it is a common technique in transfer learning.
In this post, we demonstrated how you can use Isaac Sim to train an AI driver for a simulated JetBot and transfer the skill to a real one. You can evaluate how well a trained RL model performs on the real JetBot, then use Isaac Sim to address shortcomings. Solutions include randomizing more around failed cases, using domain randomization for lighting glares, camera calibration, and so on, and retraining and redeploying. There are more things you could try to improve the result further.
Isaac Sim also provides RGB, depth, segmentation, and bounding box data. We encourage you to use this data in Isaac Sim to explore teaching your JetBot new tricks.
Special thanks to the NVIDIA Isaac Sim team and Jetson team for contributing to this post, especially Hammad Mazhar, Renato Gasoto, Cameron Upright, Chitoku Yato and John Welsh.