sac/wrappers.py

# coding=utf-8
# Copyright 2024 The Google Research Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

"""Environment wrappers. From XIRL by Zakka et al.[1]
[1]: https://github.com/google-research/google-research/tree/master/xirl
"""

import abc
import collections
import os
import time
import typing

import cv2
import gymnasium
import mujoco
from gymnasium.envs.mujoco.mujoco_rendering import OffScreenViewer
import imageio
import numpy as np
import torch
# from xirl.models import SelfSupervisedModel

TimeStep = typing.Tuple[np.ndarray, float, bool, dict]
# ModelType = SelfSupervisedModel
TensorType = torch.Tensor
DistanceFuncType = typing.Callable[[float], float]
InfoMetric = typing.Mapping[str, typing.Mapping[str, typing.Any]]


class FrameStack(gymnasium.Wrapper):
  """Stack the last k frames of the env into a flat array.

  This is useful for allowing the RL policy to infer temporal information.

  Reference: https://github.com/ikostrikov/jaxrl/
  """

  def __init__(self, env, k):
    """Constructor.

    Args:
      env: A gym env.
      k: The number of frames to stack.
    """
    super().__init__(env)

    assert isinstance(k, int), "k must be an integer."

    self._k = k
    self._frames = collections.deque([], maxlen=k)

    shp = env.observation_space.shape
    self.observation_space = gymnasium.spaces.Box(
        low=env.observation_space.low.min(),
        high=env.observation_space.high.max(),
        shape=((shp[0] * k,) + shp[1:]),
        dtype=env.observation_space.dtype,
    )

  def reset(self, seed=None, options=None):
    obs = self.env.reset()
    for _ in range(self._k):
      self._frames.append(obs)
    return self._get_obs()

  def step(self, action):
    obs, reward, terminated, truncated, info = self.env.step(action)
    self._frames.append(obs)
    return self._get_obs(), reward, terminated, truncated, info

  def _get_obs(self):
    assert len(self._frames) == self._k
    return np.concatenate(list(self._frames), axis=0)


class ActionRepeat(gymnasium.Wrapper):
  """Repeat the agent's action N times in the environment.

  Reference: https://github.com/ikostrikov/jaxrl/
  """

  def __init__(self, env, repeat):
    """Constructor.

    Args:
      env: A gym env.
      repeat: The number of times to repeat the action per single underlying env
        step.
    """
    super().__init__(env)

    assert repeat > 1, "repeat should be greater than 1."
    self._repeat = repeat

  def step(self, action):
    total_reward = 0.0
    for _ in range(self._repeat):
      obs, rew, terminated, truncated, info = self.env.step(action)
      total_reward += rew
      if terminated:
        break
    return obs, total_reward, terminated, truncated, info


class RewardScale(gymnasium.Wrapper):
  """Scale the environment reward."""

  def __init__(self, env, scale):
    """Constructor.

    Args:
      env: A gym env.
      scale: How much to scale the reward by.
    """
    super().__init__(env)

    self._scale = scale

  def step(self, action):
    obs, reward, terminated, truncated, info = self.env.step(action)
    reward *= self._scale
    return obs, reward, terminated, truncated, info


def _find_success_fn(env):
  """Return a callable that checks task success on the underlying env, if any."""

  visited = set()

  def dfs(obj, depth=0, max_depth=20):
    if obj is None or depth > max_depth or id(obj) in visited:
      return None
    visited.add(id(obj))

    if hasattr(obj, "is_success") and callable(getattr(obj, "is_success")):
      def fn():
        res = obj.is_success()
        if isinstance(res, dict):
          return bool(res.get("task", False))
        return bool(res)
      return fn

    if hasattr(obj, "_check_success") and callable(getattr(obj, "_check_success")):
      def fn():
        return bool(obj._check_success())
      return fn

    for name in ("env", "_env", "rs_env", "_rs_env", "wrapped_env", "unwrapped"):
      if hasattr(obj, name):
        child = getattr(obj, name)
        if child is not obj:
          candidate = dfs(child, depth + 1, max_depth)
          if candidate is not None:
            return candidate

    return None

  return dfs(env)


class TerminateOnSuccess(gymnasium.Wrapper):
  """Terminate the episode as soon as the underlying env reports task success."""

  def __init__(self, env):
    super().__init__(env)
    self._success_fn = _find_success_fn(env)

  def step(self, action):
    obs, rew, terminated, truncated, info = self.env.step(action)
    if not terminated and self._success_fn is not None:
      try:
        if self._success_fn():
          terminated = True
          info["success"] = True
      except Exception:
        pass
    return obs, rew, terminated, truncated, info


class EpisodeMonitor(gymnasium.Wrapper):
  """A class that computes episode metrics.

  At minimum, episode return, length and duration are computed. Additional
  metrics that are logged in the environment's info dict can be monitored by
  specifying them via `info_metrics`.

  Reference: https://github.com/ikostrikov/jaxrl/
  """

  def __init__(self, env):
    super().__init__(env)

    self._reset_stats()
    self.total_timesteps: int = 0

    self._success_fn = _find_success_fn(self.env)

  def _reset_stats(self):
    self.reward_sum: float = 0.0
    self.episode_length: int = 0
    self.start_time = time.time()

  def step(self, action):
    obs, rew, terminated, truncated, info = self.env.step(action)

    self.reward_sum += rew
    self.episode_length += 1
    self.total_timesteps += 1
    info["total"] = {"timesteps": self.total_timesteps}

    if terminated:
      info["episode"] = dict()
      info["episode"]["return"] = self.reward_sum
      info["episode"]["length"] = self.episode_length
      info["episode"]["duration"] = time.time() - self.start_time
      info["episode"]["success"] = self._final_success()

    return obs, rew, terminated, truncated, info

  def reset(self, seed=None, options=None):
    self._reset_stats()
    return self.env.reset()

  def _final_success(self) -> bool:
    """Return success for the *current* final state, or False if unavailable."""
    if self._success_fn is None:
      return False
    try:
      return bool(self._success_fn())
    except Exception:
      # Never crash training because of logging
      return False


class VideoRecorder(gymnasium.Wrapper):
  """Wrapper for rendering and saving rollouts to disk.

  Reference: https://github.com/ikostrikov/jaxrl/
  """

  def __init__(
      self,
      env,
      save_dir,
      resolution = (84, 84),
      fps = 30,
      camera = "corner2", # ['corner', 'corner2', 'corner3']
  ):
    super().__init__(env)

    self.save_dir = save_dir
    os.makedirs(save_dir, exist_ok=True)

    self.width, self.height = resolution
    self.resolution = resolution
    self.fps = fps
    self.enabled = True
    self.current_episode = 0
    self.frames = []
    self.camera = camera

  def step(self, action):
    frame = self.env.render()
      # offscreen=True, 
      # camera_name=self.camera, 
      # resolution=self.resolution
    # )
    frame = np.flipud(np.array(frame))

    if frame.shape[:2] != (self.width, self.height):
      frame = cv2.resize(
          frame,
          dsize=(self.width, self.height),
          interpolation=cv2.INTER_CUBIC,
      )
    self.frames.append(frame)
    observation, reward, terminated, truncated, info = self.env.step(action)
    if truncated or terminated:
      filename = os.path.join(self.save_dir, f"{self.current_episode}.mp4")
      imageio.mimsave(filename, self.frames, fps=self.fps)
      self.frames = []
      self.current_episode += 1
    return observation, reward, terminated, truncated, info


class EnforceMaxPathLength(gymnasium.Wrapper):
  """Enforce `done=True` when `max_path_length` is reached in MetaWorld."""
  
  def __init__(self, env):
    super().__init__(env)
    self.max_path_length = 300
    self.episode_steps = 0
  
  def reset(self, seed=None, options=None):
    self.episode_steps = 0
    return self.env.reset()

  def step(self, action):
    obs, rew, terminated, truncated, info = self.env.step(action)
    self.episode_steps += 1
    info['episode_steps'] = self.episode_steps

    if self.episode_steps >= self.max_path_length:
      terminated = True

    return obs, rew, terminated, truncated, info


class RenderWrapper(gymnasium.Wrapper):
  """
    Minimal RenderWrapper for robosuite.

    - Assumes observations contain a key like 'agentview_image'.
    - Extracts that image, optionally flips it vertically.
    - Stores the latest frame so env.render() returns an RGB array.
    - Everything else (step/reset) is passed through.
    """

  def __init__(self, env, camera_key="agentview_image", flip_vertical=True):
    super().__init__(env)
    self.camera_key = camera_key
    self.flip_vertical = flip_vertical
    self._last_frame = None

  def _extract_frame(self, obs):
    """
    obs can be:
      - a dict (robosuite obs with images)
      - already an image array (if a previous wrapper converted it)
    """
    if isinstance(obs, dict):
      img = obs[self.camera_key]  # e.g. obs["agentview_image"]
    else:
      # assume it's already an image array
      img = obs

    if self.flip_vertical:
      img = img[::-1, :, :]   # flip vertically, like your example

    self._last_frame = img
    return obs

  def reset(self, **kwargs):
    # gymnasium-style reset: (obs, info)
    result = self.env.reset(**kwargs)
    # support both gymnasium (obs, info) and gym (obs) just in case
    if isinstance(result, tuple) and len(result) == 2:
      obs, info = result
      obs = self._extract_frame(obs)
      return obs, info
    else:
      obs = result
      obs = self._extract_frame(obs)
      return obs

  def step(self, action):
    # gymnasium-style step: (obs, reward, terminated, truncated, info)
    result = self.env.step(action)
    if len(result) == 5:
      obs, reward, terminated, truncated, info = result
      obs = self._extract_frame(obs)
      return obs, reward, terminated, truncated, info
    else:
      # fallback to gym-style (obs, reward, done, info)
      obs, reward, done, info = result
      obs = self._extract_frame(obs)
      return obs, reward, done, info

  def render(self):
    """
    Return the latest camera frame as an RGB array, for VideoRecorder etc.
    """
    return self._last_frame

# ========================================= #
# Learned reward wrappers.
# ========================================= #

# Note: While the below classes provide a nice wrapper API, they are not
# efficient for training RL policies as rewards are computed individually at
# every `env.step()` and so cannot take advantage of batching on the GPU.
# For actually training policies, it is better to use the learned replay buffer
# implementations in `sac.replay_buffer.py`. These store transitions in a
# staging buffer which is forwarded as a batch through the GPU.


class LearnedVisualReward(abc.ABC, gymnasium.Wrapper):
  """Base wrapper class that replaces the env reward with a learned one.

  Subclasses should implement the `_get_reward_from_image` method.
  """

  def __init__(
      self,
      env,
      model,
      device,
      res_hw = None,
  ):
    """Constructor.

    Args:
      env: A gym env.
      model: A model that ingests RGB frames and returns embeddings. Should be a
        subclass of `xirl.models.SelfSupervisedModel`.
      device: Compute device.
      res_hw: Optional (H, W) to resize the environment image before feeding it
        to the model.
    """
    super().__init__(env)

    self._device = device
    self._model = model.to(device).eval()
    self._res_hw = res_hw

  def _to_tensor(self, x):
    x = torch.from_numpy(x).permute(2, 0, 1).float()[None, None, Ellipsis]
    # TODO(kevin): Make this more generic for other preprocessing.
    x = x / 255.0
    x = x.to(self._device)
    return x

  def _render_obs(self):
    """Render the pixels at the desired resolution."""
    # TODO(kevin): Make sure this works for mujoco envs.
    pixels = self.env.render(mode="rgb_array")
    if self._res_hw is not None:
      h, w = self._res_hw
      pixels = cv2.resize(pixels, dsize=(w, h), interpolation=cv2.INTER_CUBIC)
    return pixels

  @abc.abstractmethod
  def _get_reward_from_image(self, image):
    """Forward the pixels through the model and compute the reward."""

  def step(self, action):
    obs, env_reward, done, info = self.env.step(action)
    # We'll keep the original env reward in the info dict in case the user would
    # like to use it in conjunction with the learned reward.
    info["env_reward"] = env_reward
    pixels = self._render_obs()
    learned_reward = self._get_reward_from_image(pixels)
    return obs, learned_reward, done, info


class DistanceToGoalLearnedVisualReward(LearnedVisualReward):
  """Replace the environment reward with distances in embedding space."""

  def __init__(
      self,
      goal_emb,
      distance_scale = 1.0,
      **base_kwargs,
  ):
    """Constructor.

    Args:
      goal_emb: The goal embedding.
      distance_scale: Scales the distance from the current state embedding to
        that of the goal state. Set to `1.0` by default.
      **base_kwargs: Base keyword arguments.
    """
    super().__init__(**base_kwargs)

    self._goal_emb = np.atleast_2d(goal_emb)
    self._distance_scale = distance_scale

  def _get_reward_from_image(self, image):
    """Forward the pixels through the model and compute the reward."""
    image_tensor = self._to_tensor(image)
    emb = self._model.infer(image_tensor).numpy().embs
    dist = -1.0 * np.linalg.norm(emb - self._goal_emb)
    dist *= self._distance_scale
    return dist


class GoalClassifierLearnedVisualReward(LearnedVisualReward):
  """Replace the environment reward with the output of a goal classifier."""

  def _get_reward_from_image(self, image):
    """Forward the pixels through the model and compute the reward."""
    image_tensor = self._to_tensor(image)
    prob = torch.sigmoid(self._model.infer(image_tensor).embs)
    return prob.item()