How to construct an RLDS dataset?

[Loong1512]

I want to fine-tune the Octo model, but I don't know how to construct my own RLDS dataset. I have already built a reinforcement learning environment using dm_env, performed simulation with Isaac Gym, and generated an RLDS dataset using envlogger.

for i in range(FLAGS.num_episodes):
  timestep = env.reset()
  while True:
    # TODO: HOW TO GENERATE ACTION
    action = np.random.uniform(-3, 3, size=(9,)).astype(np.float32)
    timestep = env.step(action)
    gym_env.render()
gym_env.close()

However, for the action part, how should I control the robotic arm to generate grasping motion trajectories? Should it be learned through a reward function, controlled via a keyboard or gamepad, or through motion capture? I'm quite confused about this. Could someone tell me the general solution?