Reset vectorized environment

[giangdao1402]

Currently, i am developing an obstacle avoidance task for carter_v1. I have already attached the contact sensor to the robot. Everything goes well. I want to reset the episode each time the robot collide with obstacles, so i have implemented the logic in the _get_done function. But there is a problem, when training multiple robots (vectorized environment), if one collided and reset, other ones also reset (meaning that the _get_done of one agent affect others). Can someone give me some suggestions to solve this problem ?

[RandomOakForest]

Thanks for posting this. Could you share your _get_dones function? Chances are when you are detecting a collision you are setting all environments to a failed termination state instead of just the relevant one.

[giangdao1402]

Here is my _get_dones function

def _get_dones(self) -> tuple[torch.Tensor, torch.Tensor]:
    # Check if the task has exceeded the maximum length
    timeout = self.episode_length_buf >= self.max_episode_length - 1

    # Check crash condition
    crashed = self._compute_crash()

    # Combine both failure conditions
    task_failed = timeout | crashed  # element-wise OR for torch tensors
    return task_failed, self.task_completed