search

utiasDSL / gym-pybullet-drones

PyBullet Gymnasium environments for single and multi-agent reinforcement learning of quadcopter control
https://utiasDSL.github.io/gym-pybullet-drones/
MIT License
1.21k stars 352 forks source link

How to control quadcopter using keyboard? #109

Open HimGautam opened 2 years ago

HimGautam commented 2 years ago

Hi @JacopoPan , I wanted to do a project using deep learning. For this, I wanted offline data containing states, actions, next_states. So is there an environment in which I can directly control the quadcopter from my keyboard?

JacopoPan commented 2 years ago

Hi @HimGautam, the functionality you are looking for is not currently implemented. It is technically possible to do and not overly complicated (e.g., starting from a class like VelocityAviary that already contains a PID controller and using something like pygame). I would have one major doubt: do you really need human interaction (or could you just script the behaviour beforehand)? Human-in-the-loop data collection might be to great a bottleneck for many deep learning tasks.

HimGautam commented 2 years ago

Human in the loop is not required. It is just that a diverse dataset is needed.

JacopoPan commented 2 years ago

You could just use fly.py with a single drone (DEFAULT_NUM_DRONES = 1) and change TARGET_POS = np.zeros((NUM_WP,3)) to the trajectory you want.

HarshilNaikk commented 1 year ago

@JacopoPan Can you elaborate a bit more about using your own trajectory? Cause when I try to use my own trajectory using algorithms like RRT*, the drone just crashes every time. I see variation in its behaviour when I change the parameters "NUM_WP, simulation_freq_hq, control_freq, and aggregate_phy_steps". Can you expand a bit more about how you set and tune these parameters? I have a feeling NUM_WP is not just equal to the number of waypoints in your trajectory. It might have more to do with the actual way the waypoints are passed to the controller as well?

JacopoPan commented 1 year ago

@HarshilNaikk

editing fly.py still relies on DSLPIDControl.py whose parameters are tuned to track a trajectory of (at least) positions (and possibly velocities) that are not too far from one another.

If you are only generating sparse waypoints with a planner, you'd need to interpolate them, e.g. with a polynomial fit. Otherwise, if you have waypoints far apart, the PID controller will try to "jump" between them and lose control.