The position of rigid_body will change repeatedly between two different positions when the environment is reset

Robotics & Edge Computing Robotics - Isaac Isaac Gym
1

In the “compute _observations” function of “Franka Cabinet”, the following code can read the position of rigid_body normally

 self.franka_lfinger_pos = self.rigid_body_states[:, self.lfinger_handle][:, 0:3]
 self.franka_rfinger_pos = self.rigid_body_states[:, self.rfinger_handle][:, 0:3]

I tried to imitate FrankaCabinet, but found that with theses codes the position of rigid_body will change repeatedly between two different positions when the environment is reset(“reset_idx”).

I did not use “gym.get_rigid_transform” and “gymapi.Transform()” to declare variables. Will this have an impact?

Here is the entire code about the variable like “self.franka_lfinger_pos”:

def __init__(self, cfg, sim_device, graphics_device_id, headless):
          self.racket_pos=torch.zeros_like(self.rigid_body_states[:, self.racket_handle][:, 0:3])
          self.ball_pos=torch.zeros_like(self.rigid_body_states[:, self.racket_handle][:, 0:3])



def compute_observations(self):
    self.racket_pos = self.rigid_body_states[:, self.racket_handle][:, 0:3]
    self.ball_pos = self.rigid_body_states[:, self.ball_handle][:, 0:3]
    self.obs_buf = torch.cat((dof_pos_scaled, self.Robot_dof_vel * self.dof_vel_scale, self.racket_pos, self.ball_pos), dim=-1)


 def compute_reward(self, actions):
   self.rew_buf[:], self.reset_buf[:] = compute_PingPong_reward(
   self.reset_buf, self.progress_buf, self.actions, self.racket_pos,self.ball_pos,
   self.num_envs, self.dist_reward_scale,self.action_penalty_scale, self.max_episode_length
           )

The values I set are as follows, but I think it should be correct.

def reset_idx(self, env_ids):
ball_pos = self.ball_pos_lower_limits+torch.rand((len(env_ids),1, 3), device=self.device)*(self.ball_pos_higher_limits-self.ball_pos_lower_limits)
        self.ball_rigid_pos[env_ids] = ball_pos.squeeze(1)

    ball_state=torch.cat((ball_pos,self.default_ball_states[env_ids,...,3:]),dim=-1)
    self.ball_root_states[env_ids] = ball_state

    self.gym.set_actor_root_state_tensor(self.sim,
                                                gymtorch.unwrap_tensor(self.root_state_tensor)),


    self.gym.set_dof_position_target_tensor(self.sim,gymtorch.unwrap_tensor(self.Robot_dof_targets))

    self.gym.set_dof_state_tensor(self.sim,gymtorch.unwrap_tensor(self.dof_state))

    self.progress_buf[env_ids] = 0
    self.reset_buf[env_ids] = 0

As long as I delete these codes in “compute_observations(self)”, the environment can be reset normally.

    ​self.racket_pos = self.rigid_body_states[:, self.racket_handle][:, 0:3]
    self.ball_pos = self.rigid_body_states[:, self.ball_handle][:, 0:3]

Besides, if I use local variable , I can also reset the environment correctly.

However, with these method I can’t record the position of the balls and rackets.

1 Like
2

I can’t say for sure because I don’t have an IssacGym environment right now, but do you need the code below? Does the sample look like that?

self.racket_pos=torch.zeros_like(self.rigid_body_states[:, self.racket_handle][:, 0:3])
self.ball_pos=torch.zeros_like(self.rigid_body_states[:, self.racket_handle][:, 0:3])

2021年12月15日(水) 22:23 780927810 via NVIDIA Developer Forums <nvidia@discoursemail.com>:

3

No, in isaac gym env, these functions I wrote are used by the vec_env class. I can’t explain its specific logic, but I sincerely suggest you refer to the issac gym env code library. It is so great and Inspires me a lot of coding skills

4

It’s hard to say from the snippets you posted. What values are you setting the position buffers to when you reset?

5

The values I set are as follows, but I think it should be correct.

def reset_idx(self, env_ids):
ball_pos = self.ball_pos_lower_limits+torch.rand((len(env_ids),1, 3), device=self.device)*(self.ball_pos_higher_limits-self.ball_pos_lower_limits)
        self.ball_rigid_pos[env_ids] = ball_pos.squeeze(1)

    ball_state=torch.cat((ball_pos,self.default_ball_states[env_ids,...,3:]),dim=-1)
    self.ball_root_states[env_ids] = ball_state

    self.gym.set_actor_root_state_tensor(self.sim,
                                                gymtorch.unwrap_tensor(self.root_state_tensor)),


    self.gym.set_dof_position_target_tensor(self.sim,gymtorch.unwrap_tensor(self.Robot_dof_targets))

    self.gym.set_dof_state_tensor(self.sim,gymtorch.unwrap_tensor(self.dof_state))

    self.progress_buf[env_ids] = 0
    self.reset_buf[env_ids] = 0

As long as I delete these codes in “compute_observations(self)”, the environment can be reset normally.

    ​self.racket_pos = self.rigid_body_states[:, self.racket_handle][:, 0:3]
    self.ball_pos = self.rigid_body_states[:, self.ball_handle][:, 0:3]

Besides, if I use local variable , I can also reset the environment correctly.

However, with these method I can’t record the position of the balls and rackets.

6

In the first part of the “compute_observations” function,
Missing “self.gym.refresh_rigid_body_state_tensor (self.sim)”.
It is absolutely necessary for headless mode.
With non-headless mode, it works for some reason without it.

Even if the non-headless mode is good, the headless mode may not work, so it is necessary to check if the two match.

7

In the function “reset_idx”, to reset the position of the ball and robot, I used the following functions :

  self.gym.set_actor_root_state_tensor
  self.gym.set_dof_position_target_tensor 
  self.gym.set_dof_state_tensor

However, I solved the problem with the following functions.

  self.gym.set_actor_root_state_tensor_indexed
  self.gym.set_dof_position_target_tensor_indexed
  self.gym.set_dof_state_tensor indexed
1 Like
8

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.