Add online training with TD-MPC as proof of concept (#338)

lerobot/common/datasets/online_buffer.py

@@ -0,0 +1,384 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# 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.
+"""An online buffer for the online training loop in train.py
+
+Note to maintainers: This duplicates some logic from LeRobotDataset and EpisodeAwareSampler. We should
+consider converging to one approach. Here we have opted to use numpy.memmap to back the data buffer. It's much
+faster than using HuggingFace Datasets as there's no conversion to an intermediate non-python object. Also it
+supports in-place slicing and mutation which is very handy for a dynamic buffer.
+"""
+
+import os
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import torch
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+
+
+def _make_memmap_safe(**kwargs) -> np.memmap:
+    """Make a numpy memmap with checks on available disk space first.
+
+    Expected kwargs are: "filename", "dtype" (must by np.dtype), "mode" and "shape"
+
+    For information on dtypes:
+    https://numpy.org/doc/stable/reference/arrays.dtypes.html#arrays-dtypes-constructing
+    """
+    if kwargs["mode"].startswith("w"):
+        required_space = kwargs["dtype"].itemsize * np.prod(kwargs["shape"])  # bytes
+        stats = os.statvfs(Path(kwargs["filename"]).parent)
+        available_space = stats.f_bavail * stats.f_frsize  # bytes
+        if required_space >= available_space * 0.8:
+            raise RuntimeError(
+                f"You're about to take up {required_space} of {available_space} bytes available."
+            )
+    return np.memmap(**kwargs)
+
+
+class OnlineBuffer(torch.utils.data.Dataset):
+    """FIFO data buffer for the online training loop in train.py.
+
+    Follows the protocol of LeRobotDataset as much as is required to have it be used by the online training
+    loop in the same way that a LeRobotDataset would be used.
+
+    The underlying data structure will have data inserted in a circular fashion. Always insert after the
+    last index, and when you reach the end, wrap around to the start.
+
+    The data is stored in a numpy memmap.
+    """
+
+    NEXT_INDEX_KEY = "_next_index"
+    OCCUPANCY_MASK_KEY = "_occupancy_mask"
+    INDEX_KEY = "index"
+    FRAME_INDEX_KEY = "frame_index"
+    EPISODE_INDEX_KEY = "episode_index"
+    TIMESTAMP_KEY = "timestamp"
+    IS_PAD_POSTFIX = "_is_pad"
+
+    def __init__(
+        self,
+        write_dir: str | Path,
+        data_spec: dict[str, Any] | None,
+        buffer_capacity: int | None,
+        fps: float | None = None,
+        delta_timestamps: dict[str, list[float]] | dict[str, np.ndarray] | None = None,
+    ):
+        """
+        The online buffer can be provided from scratch or you can load an existing online buffer by passing
+        a `write_dir` associated with an existing buffer.
+
+        Args:
+            write_dir: Where to keep the numpy memmap files. One memmap file will be stored for each data key.
+                Note that if the files already exist, they are opened in read-write mode (used for training
+                resumption.)
+            data_spec: A mapping from data key to data specification, like {data_key: {"shape": tuple[int],
+                "dtype": np.dtype}}. This should include all the data that you wish to record into the buffer,
+                but note that "index", "frame_index" and "episode_index" are already accounted for by this
+                class, so you don't need to include them.
+            buffer_capacity: How many frames should be stored in the buffer as a maximum. Be aware of your
+                system's available disk space when choosing this.
+            fps: Same as the fps concept in LeRobot dataset. Here it needs to be provided for the
+                 delta_timestamps logic. You can pass None if you are not using delta_timestamps.
+            delta_timestamps: Same as the delta_timestamps concept in LeRobotDataset. This is internally
+                converted to dict[str, np.ndarray] for optimization purposes.
+
+        """
+        self.set_delta_timestamps(delta_timestamps)
+        self._fps = fps
+        # Tolerance in seconds used to discard loaded frames when their timestamps are not close enough from
+        # the requested frames. It is only used when `delta_timestamps` is provided.
+        # minus 1e-4 to account for possible numerical error
+        self.tolerance_s = 1 / self.fps - 1e-4 if fps is not None else None
+        self._buffer_capacity = buffer_capacity
+        data_spec = self._make_data_spec(data_spec, buffer_capacity)
+        Path(write_dir).mkdir(parents=True, exist_ok=True)
+        self._data = {}
+        for k, v in data_spec.items():
+            self._data[k] = _make_memmap_safe(
+                filename=Path(write_dir) / k,
+                dtype=v["dtype"] if v is not None else None,
+                mode="r+" if (Path(write_dir) / k).exists() else "w+",
+                shape=tuple(v["shape"]) if v is not None else None,
+            )
+
+    @property
+    def delta_timestamps(self) -> dict[str, np.ndarray] | None:
+        return self._delta_timestamps
+
+    def set_delta_timestamps(self, value: dict[str, list[float]] | None):
+        """Set delta_timestamps converting the values to numpy arrays.
+
+        The conversion is for an optimization in the __getitem__. The loop is much slower if the arrays
+        need to be converted into numpy arrays.
+        """
+        if value is not None:
+            self._delta_timestamps = {k: np.array(v) for k, v in value.items()}
+        else:
+            self._delta_timestamps = None
+
+    def _make_data_spec(self, data_spec: dict[str, Any], buffer_capacity: int) -> dict[str, dict[str, Any]]:
+        """Makes the data spec for np.memmap."""
+        if any(k.startswith("_") for k in data_spec):
+            raise ValueError(
+                "data_spec keys should not start with '_'. This prefix is reserved for internal logic."
+            )
+        preset_keys = {
+            OnlineBuffer.INDEX_KEY,
+            OnlineBuffer.FRAME_INDEX_KEY,
+            OnlineBuffer.EPISODE_INDEX_KEY,
+            OnlineBuffer.TIMESTAMP_KEY,
+        }
+        if len(intersection := set(data_spec).intersection(preset_keys)) > 0:
+            raise ValueError(
+                f"data_spec should not contain any of {preset_keys} as these are handled internally. "
+                f"The provided data_spec has {intersection}."
+            )
+        complete_data_spec = {
+            # _next_index will be a pointer to the next index that we should start filling from when we add
+            # more data.
+            OnlineBuffer.NEXT_INDEX_KEY: {"dtype": np.dtype("int64"), "shape": ()},
+            # Since the memmap is initialized with all-zeros, this keeps track of which indices are occupied
+            # with real data rather than the dummy initialization.
+            OnlineBuffer.OCCUPANCY_MASK_KEY: {"dtype": np.dtype("?"), "shape": (buffer_capacity,)},
+            OnlineBuffer.INDEX_KEY: {"dtype": np.dtype("int64"), "shape": (buffer_capacity,)},
+            OnlineBuffer.FRAME_INDEX_KEY: {"dtype": np.dtype("int64"), "shape": (buffer_capacity,)},
+            OnlineBuffer.EPISODE_INDEX_KEY: {"dtype": np.dtype("int64"), "shape": (buffer_capacity,)},
+            OnlineBuffer.TIMESTAMP_KEY: {"dtype": np.dtype("float64"), "shape": (buffer_capacity,)},
+        }
+        for k, v in data_spec.items():
+            complete_data_spec[k] = {"dtype": v["dtype"], "shape": (buffer_capacity, *v["shape"])}
+        return complete_data_spec
+
+    def add_data(self, data: dict[str, np.ndarray]):
+        """Add new data to the buffer, which could potentially mean shifting old data out.
+
+        The new data should contain all the frames (in order) of any number of episodes. The indices should
+        start from 0 (note to the developer: this can easily be generalized). See the `rollout` and
+        `eval_policy` functions in `eval.py` for more information on how the data is constructed.
+
+        Shift the incoming data index and episode_index to continue on from the last frame. Note that this
+        will be done in place!
+        """
+        if len(missing_keys := (set(self.data_keys).difference(set(data)))) > 0:
+            raise ValueError(f"Missing data keys: {missing_keys}")
+        new_data_length = len(data[self.data_keys[0]])
+        if not all(len(data[k]) == new_data_length for k in self.data_keys):
+            raise ValueError("All data items should have the same length")
+
+        next_index = self._data[OnlineBuffer.NEXT_INDEX_KEY]
+
+        # Sanity check to make sure that the new data indices start from 0.
+        assert data[OnlineBuffer.EPISODE_INDEX_KEY][0].item() == 0
+        assert data[OnlineBuffer.INDEX_KEY][0].item() == 0
+
+        # Shift the incoming indices if necessary.
+        if self.num_samples > 0:
+            last_episode_index = self._data[OnlineBuffer.EPISODE_INDEX_KEY][next_index - 1]
+            last_data_index = self._data[OnlineBuffer.INDEX_KEY][next_index - 1]
+            data[OnlineBuffer.EPISODE_INDEX_KEY] += last_episode_index + 1
+            data[OnlineBuffer.INDEX_KEY] += last_data_index + 1
+
+        # Insert the new data starting from next_index. It may be necessary to wrap around to the start.
+        n_surplus = max(0, new_data_length - (self._buffer_capacity - next_index))
+        for k in self.data_keys:
+            if n_surplus == 0:
+                slc = slice(next_index, next_index + new_data_length)
+                self._data[k][slc] = data[k]
+                self._data[OnlineBuffer.OCCUPANCY_MASK_KEY][slc] = True
+            else:
+                self._data[k][next_index:] = data[k][:-n_surplus]
+                self._data[OnlineBuffer.OCCUPANCY_MASK_KEY][next_index:] = True
+                self._data[k][:n_surplus] = data[k][-n_surplus:]
+        if n_surplus == 0:
+            self._data[OnlineBuffer.NEXT_INDEX_KEY] = next_index + new_data_length
+        else:
+            self._data[OnlineBuffer.NEXT_INDEX_KEY] = n_surplus
+
+    @property
+    def data_keys(self) -> list[str]:
+        keys = set(self._data)
+        keys.remove(OnlineBuffer.OCCUPANCY_MASK_KEY)
+        keys.remove(OnlineBuffer.NEXT_INDEX_KEY)
+        return sorted(keys)
+
+    @property
+    def fps(self) -> float | None:
+        return self._fps
+
+    @property
+    def num_episodes(self) -> int:
+        return len(
+            np.unique(self._data[OnlineBuffer.EPISODE_INDEX_KEY][self._data[OnlineBuffer.OCCUPANCY_MASK_KEY]])
+        )
+
+    @property
+    def num_samples(self) -> int:
+        return np.count_nonzero(self._data[OnlineBuffer.OCCUPANCY_MASK_KEY])
+
+    def __len__(self):
+        return self.num_samples
+
+    def _item_to_tensors(self, item: dict) -> dict:
+        item_ = {}
+        for k, v in item.items():
+            if isinstance(v, torch.Tensor):
+                item_[k] = v
+            elif isinstance(v, np.ndarray):
+                item_[k] = torch.from_numpy(v)
+            else:
+                item_[k] = torch.tensor(v)
+        return item_
+
+    def __getitem__(self, idx: int) -> dict[str, torch.Tensor]:
+        if idx >= len(self) or idx < -len(self):
+            raise IndexError
+
+        item = {k: v[idx] for k, v in self._data.items() if not k.startswith("_")}
+
+        if self.delta_timestamps is None:
+            return self._item_to_tensors(item)
+
+        episode_index = item[OnlineBuffer.EPISODE_INDEX_KEY]
+        current_ts = item[OnlineBuffer.TIMESTAMP_KEY]
+        episode_data_indices = np.where(
+            np.bitwise_and(
+                self._data[OnlineBuffer.EPISODE_INDEX_KEY] == episode_index,
+                self._data[OnlineBuffer.OCCUPANCY_MASK_KEY],
+            )
+        )[0]
+        episode_timestamps = self._data[OnlineBuffer.TIMESTAMP_KEY][episode_data_indices]
+
+        for data_key in self.delta_timestamps:
+            # Note: The logic in this loop is copied from `load_previous_and_future_frames`.
+            # Get timestamps used as query to retrieve data of previous/future frames.
+            query_ts = current_ts + self.delta_timestamps[data_key]
+
+            # Compute distances between each query timestamp and all timestamps of all the frames belonging to
+            # the episode.
+            dist = np.abs(query_ts[:, None] - episode_timestamps[None, :])
+            argmin_ = np.argmin(dist, axis=1)
+            min_ = dist[np.arange(dist.shape[0]), argmin_]
+
+            is_pad = min_ > self.tolerance_s
+
+            # Check violated query timestamps are all outside the episode range.
+            assert (
+                (query_ts[is_pad] < episode_timestamps[0]) | (episode_timestamps[-1] < query_ts[is_pad])
+            ).all(), (
+                f"One or several timestamps unexpectedly violate the tolerance ({min_} > {self.tolerance_s=}"
+                ") inside the episode range."
+            )
+
+            # Load frames for this data key.
+            item[data_key] = self._data[data_key][episode_data_indices[argmin_]]
+
+            item[f"{data_key}{OnlineBuffer.IS_PAD_POSTFIX}"] = is_pad
+
+        return self._item_to_tensors(item)
+
+    def get_data_by_key(self, key: str) -> torch.Tensor:
+        """Returns all data for a given data key as a Tensor."""
+        return torch.from_numpy(self._data[key][self._data[OnlineBuffer.OCCUPANCY_MASK_KEY]])
+
+
+def compute_sampler_weights(
+    offline_dataset: LeRobotDataset,
+    offline_drop_n_last_frames: int = 0,
+    online_dataset: OnlineBuffer | None = None,
+    online_sampling_ratio: float | None = None,
+    online_drop_n_last_frames: int = 0,
+) -> torch.Tensor:
+    """Compute the sampling weights for the online training dataloader in train.py.
+
+    Args:
+        offline_dataset: The LeRobotDataset used for offline pre-training.
+        online_drop_n_last_frames: Number of frames to drop from the end of each offline dataset episode.
+        online_dataset: The OnlineBuffer used in online training.
+        online_sampling_ratio: The proportion of data that should be sampled from the online dataset. If an
+            online dataset is provided, this value must also be provided.
+        online_drop_n_first_frames: See `offline_drop_n_last_frames`. This is the same, but for the online
+            dataset.
+    Returns:
+        Tensor of weights for [offline_dataset; online_dataset], normalized to 1.
+
+    Notes to maintainers:
+        - This duplicates some logic from EpisodeAwareSampler. We should consider converging to one approach.
+        - When used with `torch.utils.data.WeightedRandomSampler`, it could completely replace
+          `EpisodeAwareSampler` as the online dataset related arguments are optional. The only missing feature
+          is the ability to turn shuffling off.
+        - Options `drop_first_n_frames` and `episode_indices_to_use` can be added easily. They were not
+          included here to avoid adding complexity.
+    """
+    if len(offline_dataset) == 0 and (online_dataset is None or len(online_dataset) == 0):
+        raise ValueError("At least one of `offline_dataset` or `online_dataset` should be contain data.")
+    if (online_dataset is None) ^ (online_sampling_ratio is None):
+        raise ValueError(
+            "`online_dataset` and `online_sampling_ratio` must be provided together or not at all."
+        )
+    offline_sampling_ratio = 0 if online_sampling_ratio is None else 1 - online_sampling_ratio
+
+    weights = []
+
+    if len(offline_dataset) > 0:
+        offline_data_mask_indices = []
+        for start_index, end_index in zip(
+            offline_dataset.episode_data_index["from"],
+            offline_dataset.episode_data_index["to"],
+            strict=True,
+        ):
+            offline_data_mask_indices.extend(
+                range(start_index.item(), end_index.item() - offline_drop_n_last_frames)
+            )
+        offline_data_mask = torch.zeros(len(offline_dataset), dtype=torch.bool)
+        offline_data_mask[torch.tensor(offline_data_mask_indices)] = True
+        weights.append(
+            torch.full(
+                size=(len(offline_dataset),),
+                fill_value=offline_sampling_ratio / offline_data_mask.sum(),
+            )
+            * offline_data_mask
+        )
+
+    if online_dataset is not None and len(online_dataset) > 0:
+        online_data_mask_indices = []
+        episode_indices = online_dataset.get_data_by_key("episode_index")
+        for episode_idx in torch.unique(episode_indices):
+            where_episode = torch.where(episode_indices == episode_idx)
+            start_index = where_episode[0][0]
+            end_index = where_episode[0][-1] + 1
+            online_data_mask_indices.extend(
+                range(start_index.item(), end_index.item() - online_drop_n_last_frames)
+            )
+        online_data_mask = torch.zeros(len(online_dataset), dtype=torch.bool)
+        online_data_mask[torch.tensor(online_data_mask_indices)] = True
+        weights.append(
+            torch.full(
+                size=(len(online_dataset),),
+                fill_value=online_sampling_ratio / online_data_mask.sum(),
+            )
+            * online_data_mask
+        )
+
+    weights = torch.cat(weights)
+
+    if weights.sum() == 0:
+        weights += 1 / len(weights)
+    else:
+        weights /= weights.sum()
+
+    return weights

lerobot/common/policies/tdmpc/configuration_tdmpc.py

@@ -25,12 +25,16 @@ class TDMPCConfig:
     camera observations.
 
     The parameters you will most likely need to change are the ones which depend on the environment / sensors.
-    Those are: `input_shapes`, `output_shapes`, and perhaps `max_random_shift`.
+    Those are: `input_shapes`, `output_shapes`, and perhaps `max_random_shift_ratio`.
 
     Args:
         n_action_repeats: The number of times to repeat the action returned by the planning. (hint: Google
             action repeats in Q-learning or ask your favorite chatbot)
         horizon: Horizon for model predictive control.
+        n_action_steps: Number of action steps to take from the plan given by model predictive control. This
+            is an alternative to using action repeats. If this is set to more than 1, then we require
+            `n_action_repeats == 1`, `use_mpc == True` and `n_action_steps <= horizon`. Note that this
+            approach of using multiple steps from the plan is not in the original implementation.
         input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
             the input data name, and the value is a list indicating the dimensions of the corresponding data.
             For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
@@ -100,6 +104,7 @@ class TDMPCConfig:
     # Input / output structure.
     n_action_repeats: int = 2
     horizon: int = 5
+    n_action_steps: int = 1
 
     input_shapes: dict[str, list[int]] = field(
         default_factory=lambda: {
@@ -158,17 +163,18 @@ class TDMPCConfig:
         """Input validation (not exhaustive)."""
         # There should only be one image key.
         image_keys = {k for k in self.input_shapes if k.startswith("observation.image")}
-        if len(image_keys) != 1:
+        if len(image_keys) > 1:
             raise ValueError(
-                f"{self.__class__.__name__} only handles one image for now. Got image keys {image_keys}."
-            )
-        image_key = next(iter(image_keys))
-        if self.input_shapes[image_key][-2] != self.input_shapes[image_key][-1]:
-            # TODO(alexander-soare): This limitation is solely because of code in the random shift
-            # augmentation. It should be able to be removed.
-            raise ValueError(
-                f"Only square images are handled now. Got image shape {self.input_shapes[image_key]}."
+                f"{self.__class__.__name__} handles at most one image for now. Got image keys {image_keys}."
             )
+        if len(image_keys) > 0:
+            image_key = next(iter(image_keys))
+            if self.input_shapes[image_key][-2] != self.input_shapes[image_key][-1]:
+                # TODO(alexander-soare): This limitation is solely because of code in the random shift
+                # augmentation. It should be able to be removed.
+                raise ValueError(
+                    f"Only square images are handled now. Got image shape {self.input_shapes[image_key]}."
+                )
         if self.n_gaussian_samples <= 0:
             raise ValueError(
                 f"The number of guassian samples for CEM should be non-zero. Got `{self.n_gaussian_samples=}`"
@@ -179,3 +185,12 @@ class TDMPCConfig:
                 f"advised that you stick with the default. See {self.__class__.__name__} docstring for more "
                 "information."
             )
+        if self.n_action_steps > 1:
+            if self.n_action_repeats != 1:
+                raise ValueError(
+                    "If `n_action_steps > 1`, `n_action_repeats` must be left to its default value of 1."
+                )
+            if not self.use_mpc:
+                raise ValueError("If `n_action_steps > 1`, `use_mpc` must be set to `True`.")
+            if self.n_action_steps > self.horizon:
+                raise ValueError("`n_action_steps` must be less than or equal to `horizon`.")

lerobot/common/policies/tdmpc/modeling_tdmpc.py

@@ -19,14 +19,10 @@
 The comments in this code may sometimes refer to these references:
     TD-MPC paper: Temporal Difference Learning for Model Predictive Control (https://arxiv.org/abs/2203.04955)
     FOWM paper: Finetuning Offline World Models in the Real World (https://arxiv.org/abs/2310.16029)
-
-TODO(alexander-soare): Make rollout work for batch sizes larger than 1.
-TODO(alexander-soare): Use batch-first throughout.
 """
 
 # ruff: noqa: N806
 
-import logging
 from collections import deque
 from copy import deepcopy
 from functools import partial
@@ -56,9 +52,11 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
             process communication to use the xarm environment from FOWM. This is because our xarm
             environment uses newer dependencies and does not match the environment in FOWM. See
             https://github.com/huggingface/lerobot/pull/103 for implementation details.
-        - We have NOT checked that training on LeRobot reproduces SOTA results. This is a TODO.
+        - We have NOT checked that training on LeRobot reproduces the results from FOWM.
+        - Nevertheless, we have verified that we can train TD-MPC for PushT. See
+          `lerobot/configs/policy/tdmpc_pusht_keypoints.yaml`.
         - Our current xarm datasets were generated using the environment from FOWM. Therefore they do not
-            match our xarm environment.
+          match our xarm environment.
     """
 
     name = "tdmpc"
@@ -74,22 +72,6 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
                 that they will be passed with a call to `load_state_dict` before the policy is used.
         """
         super().__init__()
-        logging.warning(
-            """
-            Please note several warnings for this policy.
-
-            - Evaluation of pretrained weights created with the original FOWM code
-              (https://github.com/fyhMer/fowm) works as expected. To be precise: we trained and evaluated a
-              model with the FOWM code for the xarm_lift_medium_replay dataset. We ported the weights across
-              to LeRobot, and were able to evaluate with the same success metric. BUT, we had to use inter-
-              process communication to use the xarm environment from FOWM. This is because our xarm
-              environment uses newer dependencies and does not match the environment in FOWM. See
-              https://github.com/huggingface/lerobot/pull/103 for implementation details.
-            - We have NOT checked that training on LeRobot reproduces SOTA results. This is a TODO.
-            - Our current xarm datasets were generated using the environment from FOWM. Therefore they do not
-              match our xarm environment.
-            """
-        )
 
         if config is None:
             config = TDMPCConfig()
@@ -114,8 +96,14 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
 
         image_keys = [k for k in config.input_shapes if k.startswith("observation.image")]
         # Note: This check is covered in the post-init of the config but have a sanity check just in case.
-        assert len(image_keys) == 1
-        self.input_image_key = image_keys[0]
+        self._use_image = False
+        self._use_env_state = False
+        if len(image_keys) > 0:
+            assert len(image_keys) == 1
+            self._use_image = True
+            self.input_image_key = image_keys[0]
+        if "observation.environment_state" in config.input_shapes:
+            self._use_env_state = True
 
         self.reset()
 
@@ -125,10 +113,13 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
         called on `env.reset()`
         """
         self._queues = {
-            "observation.image": deque(maxlen=1),
             "observation.state": deque(maxlen=1),
-            "action": deque(maxlen=self.config.n_action_repeats),
+            "action": deque(maxlen=max(self.config.n_action_steps, self.config.n_action_repeats)),
         }
+        if self._use_image:
+            self._queues["observation.image"] = deque(maxlen=1)
+        if self._use_env_state:
+            self._queues["observation.environment_state"] = deque(maxlen=1)
         # Previous mean obtained from the cross-entropy method (CEM) used during MPC. It is used to warm start
         # CEM for the next step.
         self._prev_mean: torch.Tensor | None = None
@@ -137,8 +128,9 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
     def select_action(self, batch: dict[str, Tensor]) -> Tensor:
         """Select a single action given environment observations."""
         batch = self.normalize_inputs(batch)
-        batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
-        batch["observation.image"] = batch[self.input_image_key]
+        if self._use_image:
+            batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
+            batch["observation.image"] = batch[self.input_image_key]
 
         self._queues = populate_queues(self._queues, batch)
 
@@ -152,49 +144,57 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
                 batch[key] = batch[key][:, 0]
 
             # NOTE: Order of observations matters here.
-            z = self.model.encode({k: batch[k] for k in ["observation.image", "observation.state"]})
-            if self.config.use_mpc:
-                batch_size = batch["observation.image"].shape[0]
-                # Batch processing is not handled in MPC mode, so process the batch in a loop.
-                action = []  # will be a batch of actions for one step
-                for i in range(batch_size):
-                    # Note: self.plan does not handle batches, hence the squeeze.
-                    action.append(self.plan(z[i]))
-                action = torch.stack(action)
+            encode_keys = []
+            if self._use_image:
+                encode_keys.append("observation.image")
+            if self._use_env_state:
+                encode_keys.append("observation.environment_state")
+            encode_keys.append("observation.state")
+            z = self.model.encode({k: batch[k] for k in encode_keys})
+            if self.config.use_mpc:  # noqa: SIM108
+                actions = self.plan(z)  # (horizon, batch, action_dim)
             else:
-                # Plan with the policy (π) alone.
-                action = self.model.pi(z)
+                # Plan with the policy (π) alone. This always returns one action so unsqueeze to get a
+                # sequence dimension like in the MPC branch.
+                actions = self.model.pi(z).unsqueeze(0)
 
-            self.unnormalize_outputs({"action": action})["action"]
+            actions = torch.clamp(actions, -1, +1)
 
-            for _ in range(self.config.n_action_repeats):
-                self._queues["action"].append(action)
+            actions = self.unnormalize_outputs({"action": actions})["action"]
+
+            if self.config.n_action_repeats > 1:
+                for _ in range(self.config.n_action_repeats):
+                    self._queues["action"].append(actions[0])
+            else:
+                # Action queue is (n_action_steps, batch_size, action_dim), so we transpose the action.
+                self._queues["action"].extend(actions[: self.config.n_action_steps])
 
         action = self._queues["action"].popleft()
-        return torch.clamp(action, -1, 1)
+        return action
 
     @torch.no_grad()
     def plan(self, z: Tensor) -> Tensor:
-        """Plan next action using TD-MPC inference.
+        """Plan sequence of actions using TD-MPC inference.
 
         Args:
-            z: (latent_dim,) tensor for the initial state.
+            z: (batch, latent_dim,) tensor for the initial state.
         Returns:
-            (action_dim,) tensor for the next action.
-
-        TODO(alexander-soare) Extend this to be able to work with batches.
+            (horizon, batch, action_dim,) tensor for the planned trajectory of actions.
         """
         device = get_device_from_parameters(self)
 
+        batch_size = z.shape[0]
+
         # Sample Nπ trajectories from the policy.
         pi_actions = torch.empty(
             self.config.horizon,
             self.config.n_pi_samples,
+            batch_size,
             self.config.output_shapes["action"][0],
             device=device,
         )
         if self.config.n_pi_samples > 0:
-            _z = einops.repeat(z, "d -> n d", n=self.config.n_pi_samples)
+            _z = einops.repeat(z, "b d -> n b d", n=self.config.n_pi_samples)
             for t in range(self.config.horizon):
                 # Note: Adding a small amount of noise here doesn't hurt during inference and may even be
                 # helpful for CEM.
@@ -203,12 +203,14 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
 
         # In the CEM loop we will need this for a call to estimate_value with the gaussian sampled
         # trajectories.
-        z = einops.repeat(z, "d -> n d", n=self.config.n_gaussian_samples + self.config.n_pi_samples)
+        z = einops.repeat(z, "b d -> n b d", n=self.config.n_gaussian_samples + self.config.n_pi_samples)
 
         # Model Predictive Path Integral (MPPI) with the cross-entropy method (CEM) as the optimization
         # algorithm.
         # The initial mean and standard deviation for the cross-entropy method (CEM).
-        mean = torch.zeros(self.config.horizon, self.config.output_shapes["action"][0], device=device)
+        mean = torch.zeros(
+            self.config.horizon, batch_size, self.config.output_shapes["action"][0], device=device
+        )
         # Maybe warm start CEM with the mean from the previous step.
         if self._prev_mean is not None:
             mean[:-1] = self._prev_mean[1:]
@@ -219,6 +221,7 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
             std_normal_noise = torch.randn(
                 self.config.horizon,
                 self.config.n_gaussian_samples,
+                batch_size,
                 self.config.output_shapes["action"][0],
                 device=std.device,
             )
@@ -227,21 +230,24 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
             # Compute elite actions.
             actions = torch.cat([gaussian_actions, pi_actions], dim=1)
             value = self.estimate_value(z, actions).nan_to_num_(0)
-            elite_idxs = torch.topk(value, self.config.n_elites, dim=0).indices
-            elite_value, elite_actions = value[elite_idxs], actions[:, elite_idxs]
+            elite_idxs = torch.topk(value, self.config.n_elites, dim=0).indices  # (n_elites, batch)
+            elite_value = value.take_along_dim(elite_idxs, dim=0)  # (n_elites, batch)
+            # (horizon, n_elites, batch, action_dim)
+            elite_actions = actions.take_along_dim(einops.rearrange(elite_idxs, "n b -> 1 n b 1"), dim=1)
 
-            # Update guassian PDF parameters to be the (weighted) mean and standard deviation of the elites.
-            max_value = elite_value.max(0)[0]
+            # Update gaussian PDF parameters to be the (weighted) mean and standard deviation of the elites.
+            max_value = elite_value.max(0, keepdim=True)[0]  # (1, batch)
             # The weighting is a softmax over trajectory values. Note that this is not the same as the usage
             # of Ω in eqn 4 of the TD-MPC paper. Instead it is the normalized version of it: s = Ω/ΣΩ. This
             # makes the equations: μ = Σ(s⋅Γ), σ = Σ(s⋅(Γ-μ)²).
             score = torch.exp(self.config.elite_weighting_temperature * (elite_value - max_value))
-            score /= score.sum()
-            _mean = torch.sum(einops.rearrange(score, "n -> n 1") * elite_actions, dim=1)
+            score /= score.sum(axis=0, keepdim=True)
+            # (horizon, batch, action_dim)
+            _mean = torch.sum(einops.rearrange(score, "n b -> n b 1") * elite_actions, dim=1)
             _std = torch.sqrt(
                 torch.sum(
-                    einops.rearrange(score, "n -> n 1")
-                    * (elite_actions - einops.rearrange(_mean, "h d -> h 1 d")) ** 2,
+                    einops.rearrange(score, "n b -> n b 1")
+                    * (elite_actions - einops.rearrange(_mean, "h b d -> h 1 b d")) ** 2,
                     dim=1,
                 )
             )
@@ -256,11 +262,9 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
 
         # Randomly select one of the elite actions from the last iteration of MPPI/CEM using the softmax
         # scores from the last iteration.
-        actions = elite_actions[:, torch.multinomial(score, 1).item()]
+        actions = elite_actions[:, torch.multinomial(score.T, 1).squeeze(), torch.arange(batch_size)]
 
-        # Select only the first action
-        action = actions[0]
-        return action
+        return actions
 
     @torch.no_grad()
     def estimate_value(self, z: Tensor, actions: Tensor):
@@ -312,13 +316,17 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
             G -= running_discount * self.config.uncertainty_regularizer_coeff * terminal_values.std(0)
         return G
 
-    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
-        """Run the batch through the model and compute the loss."""
+    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor | float]:
+        """Run the batch through the model and compute the loss.
+
+        Returns a dictionary with loss as a tensor, and other information as native floats.
+        """
         device = get_device_from_parameters(self)
 
         batch = self.normalize_inputs(batch)
-        batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
-        batch["observation.image"] = batch[self.input_image_key]
+        if self._use_image:
+            batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
+            batch["observation.image"] = batch[self.input_image_key]
         batch = self.normalize_targets(batch)
 
         info = {}
@@ -328,12 +336,12 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
             if batch[key].ndim > 1:
                 batch[key] = batch[key].transpose(1, 0)
 
-        action = batch["action"]  # (t, b)
-        reward = batch["next.reward"]  # (t,)
+        action = batch["action"]  # (t, b, action_dim)
+        reward = batch["next.reward"]  # (t, b)
         observations = {k: v for k, v in batch.items() if k.startswith("observation.")}
 
         # Apply random image augmentations.
-        if self.config.max_random_shift_ratio > 0:
+        if self._use_image and self.config.max_random_shift_ratio > 0:
             observations["observation.image"] = flatten_forward_unflatten(
                 partial(random_shifts_aug, max_random_shift_ratio=self.config.max_random_shift_ratio),
                 observations["observation.image"],
@@ -345,7 +353,9 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
         for k in observations:
             current_observation[k] = observations[k][0]
             next_observations[k] = observations[k][1:]
-        horizon = next_observations["observation.image"].shape[0]
+        horizon, batch_size = next_observations[
+            "observation.image" if self._use_image else "observation.environment_state"
+        ].shape[:2]
 
         # Run latent rollout using the latent dynamics model and policy model.
         # Note this has shape `horizon+1` because there are `horizon` actions and a current `z`. Each action
@@ -415,7 +425,8 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
         # Compute state-action value loss (TD loss) for all of the Q functions in the ensemble.
         q_value_loss = (
             (
-                F.mse_loss(
+                temporal_loss_coeffs
+                * F.mse_loss(
                     q_preds_ensemble,
                     einops.repeat(q_targets, "t b -> e t b", e=q_preds_ensemble.shape[0]),
                     reduction="none",
@@ -464,10 +475,11 @@ class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
         action_preds = self.model.pi(z_preds[:-1])  # (t, b, a)
         # Calculate the MSE between the actions and the action predictions.
         # Note: FOWM's original code calculates the log probability (wrt to a unit standard deviation
-        # gaussian) and sums over the action dimension. Computing the log probability amounts to multiplying
-        # the MSE by 0.5 and adding a constant offset (the log(2*pi) term) . Here we drop the constant offset
-        # as it doesn't change the optimization step, and we drop the 0.5 as we instead make a configuration
-        # parameter for it (see below where we compute the total loss).
+        # gaussian) and sums over the action dimension. Computing the (negative) log probability amounts to
+        # multiplying the MSE by 0.5 and adding a constant offset (the log(2*pi)/2 term, times the action
+        # dimension). Here we drop the constant offset as it doesn't change the optimization step, and we drop
+        # the 0.5 as we instead make a configuration parameter for it (see below where we compute the total
+        # loss).
         mse = F.mse_loss(action_preds, action, reduction="none").sum(-1)  # (t, b)
         # NOTE: The original implementation does not take the sum over the temporal dimension like with the
         # other losses.
@@ -728,6 +740,16 @@ class TDMPCObservationEncoder(nn.Module):
                 nn.LayerNorm(config.latent_dim),
                 nn.Sigmoid(),
             )
+        if "observation.environment_state" in config.input_shapes:
+            self.env_state_enc_layers = nn.Sequential(
+                nn.Linear(
+                    config.input_shapes["observation.environment_state"][0], config.state_encoder_hidden_dim
+                ),
+                nn.ELU(),
+                nn.Linear(config.state_encoder_hidden_dim, config.latent_dim),
+                nn.LayerNorm(config.latent_dim),
+                nn.Sigmoid(),
+            )
 
     def forward(self, obs_dict: dict[str, Tensor]) -> Tensor:
         """Encode the image and/or state vector.
@@ -736,8 +758,11 @@ class TDMPCObservationEncoder(nn.Module):
         over all features.
         """
         feat = []
+        # NOTE: Order of observations matters here.
         if "observation.image" in self.config.input_shapes:
             feat.append(flatten_forward_unflatten(self.image_enc_layers, obs_dict["observation.image"]))
+        if "observation.environment_state" in self.config.input_shapes:
+            feat.append(self.env_state_enc_layers(obs_dict["observation.environment_state"]))
         if "observation.state" in self.config.input_shapes:
             feat.append(self.state_enc_layers(obs_dict["observation.state"]))
         return torch.stack(feat, dim=0).mean(0)