One of the nature of Eureka: The reward components should be exposed so that the LLM can continuously observe them to modify their parameters or content or decide on discard. But this function is lost during the transfer to the new version.
This is a bug fixing crossing ARD and ard_isaaclab_tasks:
ard_isaaclab_tasks part:
(1) Delete the ard_workspace and all the reward functions inside _get_reward, recording all the offical reward function into ont files.
(2) we build a computer_reward function inside the _get_reward, which should be as the "workspace" been modify by the ARD, aligning with eureka's paper and code, two outputs should be gen by the LLM, the reward_score and a reward components that can be self.extras["log"] to recorded into tensorboard.
ARD part:
prompt should be change as the new workspace is compute_reward
those scripts to navigate/capture/modifying the reward function
One of the nature of Eureka: The reward components should be exposed so that the LLM can continuously observe them to modify their parameters or content or decide on discard. But this function is lost during the transfer to the new version.
This is a bug fixing crossing ARD and ard_isaaclab_tasks:
ard_isaaclab_tasks part:
(1) Delete the ard_workspace and all the reward functions inside _get_reward, recording all the offical reward function into ont files.
(2) we build a computer_reward function inside the _get_reward, which should be as the "workspace" been modify by the ARD, aligning with eureka's paper and code, two outputs should be gen by the LLM, the reward_score and a reward components that can be self.extras["log"] to recorded into tensorboard.
ARD part:
prompt should be change as the new workspace is compute_reward
those scripts to navigate/capture/modifying the reward function