Sim2real issue: Robot fails to move toward target point and moves in irrelevant directions

[jieyuLiu529]

Problem Overview
During the sim2real experiment, I've encountered an issue where the model fails to correctly plan a path towards the target point. This problem isn't about poor obstacle avoidance. Instead, the robot controlled by the model doesn't move in the direction of the target point and may even move in a completely unrelated direction.

Detailed Description
Environment and Configuration Information
Operating System: Ubuntu 20.04
ROS Version: ROS Melodic
Python Version: Python 3.8
Code Repository Version: I'm using the link is [https://github.com/Shuijing725/CrowdNav_Sim2Real_Turtlebo]

Actual Phenomenon
After the robot starts, it doesn't move towards the input target point. Instead, it moves in a direction completely unrelated to the target. For example, when I input the target point coordinates as (x: 0, y: 6) the robot moves in (-4,-2)
During the movement, it seems that the robot doesn't plan a path based on the target point's position but continues to move in a fixed wrong direction.

Troubleshooting Attempts
I've tried the following solutions, but the problem still persists:
    Test in simulator , SUCCESS rate is over 90%.
    Checked if the input of the target point is correct and confirmed that the input coordinate values are error - free.
    Restarted the ROS nodes and the robot multiple times, but the result remains the same.
    Checked if the sensor data is normal, such as whether the messages on the /odom topic can be received normally.

Expected Help
I hope the author can help analyze the possible causes of this problem, such as whether there are errors in the code logic or if the configuration file needs adjustment. Additionally, I'd appreciate some suggestions for further troubleshooting.

[Shuijing725]

Did you open #6? If so, please avoid creating repetitive issues.

[jieyuLiu529]

Yes, I have already opened #6 , but I don't think they are the same issue. Their issue seems to lie in errors of the dynamic model, whereas mine appears to be in errors of location and path planning.

[Shuijing725]

I can think of 2 simple tests that you can do:

• Check no human scenarios: In config.py in your checkpoint's folder, set sim2real.use_dummy_detect=True so that the input state will contain no humans. Then check if the robot moves in the right direction. This test checks if the robot can reach goals without considering any collision avoidance.
• Check the robot dynamics gap: You can record some episodes in simulation by setting sim2real.record = True, then load the recorded trajectory in real by setting sim2real.load_act = True. The path for recording and loading can be changed in recorder.py. If the robot cannot reproduce the trajectory, then the robot dynamics has a sim2real gap.