train time

[ZhuPiter-1mol]

Hello！May i ask how long did it take you to train TD3? I am training the td3 in a 4090 GPU，but after 6 hours it just started evaluating the first epoch.I just change the robot's radius to 0.6.

[reiniscimurs]

Hi, that seems to be too much time before the evaluation starts. It should take max 105 minutes between evaluations if default parameters are used and it runs realtime. Change to 0.6 meters of robot radius does not have any effect in my setup. In 6 hours I would expect you to have already trained the model so something seems to be wrong here. However, I cannot reproduce any issue on my side so please provide any information that you can.

[ZhuPiter-1mol]

Hi, that seems to be too much time before the evaluation starts. It should take max 105 minutes between evaluations if default parameters are used and it runs realtime. Change to 0.6 meters of robot radius does not have any effect in my setup. In 6 hours I would expect you to have already trained the model so something seems to be wrong here. However, I cannot reproduce any issue on my side so please provide any information that you can.

Thanks for your reply!!I have another question: current robots don't account for motion constraints. If I want to add a curvature constraint（such as I don't want it to spin in circles）during evaluation, how should I do about it?

[reiniscimurs]

By default, a differential drive robot is used in the simulation. These robots do not have a curvature constraint. You might want to implement your simulation with an Ackermann model instead if you require some constraint (see https://ir-sim.readthedocs.io/en/stable/usage/configure_robots_obstacles.html)

[ZhuPiter-1mol]

By default, a differential drive robot is used in the simulation. These robots do not have a curvature constraint. You might want to implement your simulation with an Ackermann model instead if you require some constraint (see https://ir-sim.readthedocs.io/en/stable/usage/configure_robots_obstacles.html)

Thanks!When i use ackerman model, could i set the robot's shape to circle? And if I change the state dimension to 4？

[reiniscimurs]

You can try it out. These changes are quick and simple but I think Ackermann still uses 2 dimension inputs.

[ZhuPiter-1mol]

You can try it out. These changes are quick and simple but I think Ackermann still uses 2 dimension inputs.

OK, thanks!! I've made an attempt. And how can I control the obstacle avoidance distance of the strategy? If I want the robot to keep a greater distance from obstacles when avoiding them, what should I do?

[reiniscimurs]

This lambda function gives a negative reward for a distance close to an obstacle:

DRL-robot-navigation-IR-SIM/robot_nav/SIM_ENV/sim.py

Line 142 in bffc661

r3 = lambda x: 1.35 - x if x < 1.35 else 0.0

That is one way to do it and you can try doing something similar.