opentau.policies.value.reward

Reward calculation utilities for Value Policy training.

This module contains functions to calculate returns and discretize them into bins for value function training and advantage calculation.

Functions

calculate_n_step_return(success, ...[, c_neg])

Defines sparse Reward function for the pi0.6 policy to calculate advantage.

calculate_return_bins_with_equal_width(...)

Defines sparse Reward function for the pi0.6 policy to train value function network.

opentau.policies.value.reward.calculate_n_step_return(success: bool, n_steps_look_ahead: int, episode_end_idx: int, reward_normalizer: int, current_idx: int, c_neg: float = -100.0) float[source]

Defines sparse Reward function for the pi0.6 policy to calculate advantage.

Parameters:
  • success – Defines if the episode was successful or failed.

  • n_steps_look_ahead – Number of steps to look ahead for calculating reward.

  • episode_end_idx – Index of the end of the episode.

  • reward_normalizer – Maximum length of the episode for normalization.

  • current_idx – Current index of the episode.

  • c_neg – Negative reward for failed episodes. Defaults to -100.0.

Returns:

The normalized continuous reward for the n-step lookahead.

Return type:

float

opentau.policies.value.reward.calculate_return_bins_with_equal_width(success: bool, b: int, episode_end_idx: int, reward_normalizer: int, current_idx: int, c_neg: float = -100.0) tuple[int, float][source]

Defines sparse Reward function for the pi0.6 policy to train value function network.

Parameters:
  • success – Defines if the episode was successful or failed.

  • b – Number of bins to discretize the reward into, including the special bin 0.

  • episode_end_idx – Index of the end of the episode, exclusive to the last step.

  • reward_normalizer – Maximum length of the episode for normalization.

  • current_idx – Current index of the episode.

  • c_neg – Negative reward for failed episodes. Defaults to -100.0.

Returns:

A tuple containing:
  • bin_idx: The index of the reward bin.

  • return_normalized: The normalized return value in range [-1, 0].

Return type:

tuple[int, float]