Skip to content

Utils

robot_nav.utils

Pretraining

Handles loading of offline experience data and pretraining of a reinforcement learning model.

Attributes:

Name Type Description
file_names List[str]

List of YAML files containing pre-recorded environment samples.
model object

The model with prepare_state and train methods.
replay_buffer object

The buffer used to store experiences for training.
reward_function callable

Function to compute the reward from the environment state.
Source code in robot_nav/utils.py 
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
class Pretraining:
    """
    Handles loading of offline experience data and pretraining of a reinforcement learning model.

    Attributes:
        file_names (List[str]): List of YAML files containing pre-recorded environment samples.
        model (object): The model with `prepare_state` and `train` methods.
        replay_buffer (object): The buffer used to store experiences for training.
        reward_function (callable): Function to compute the reward from the environment state.
    """

    def __init__(
        self,
        file_names: List[str],
        model: object,
        replay_buffer: object,
        reward_function,
    ):
        self.file_names = file_names
        self.model = model
        self.replay_buffer = replay_buffer
        self.reward_function = reward_function

    def load_buffer(self):
        """
        Load samples from the specified files and populate the replay buffer.

        Returns:
            object: The populated replay buffer.
        """
        for file_name in self.file_names:
            print("Loading file: ", file_name)
            with open(file_name, "r") as file:
                samples = yaml.full_load(file)
                for i in tqdm(range(1, len(samples) - 1)):
                    sample = samples[i]
                    latest_scan = sample["latest_scan"]
                    distance = sample["distance"]
                    cos = sample["cos"]
                    sin = sample["sin"]
                    collision = sample["collision"]
                    goal = sample["goal"]
                    action = sample["action"]

                    state, terminal = self.model.prepare_state(
                        latest_scan, distance, cos, sin, collision, goal, action
                    )

                    if terminal:
                        continue

                    next_sample = samples[i + 1]
                    next_latest_scan = next_sample["latest_scan"]
                    next_distance = next_sample["distance"]
                    next_cos = next_sample["cos"]
                    next_sin = next_sample["sin"]
                    next_collision = next_sample["collision"]
                    next_goal = next_sample["goal"]
                    next_action = next_sample["action"]
                    next_state, next_terminal = self.model.prepare_state(
                        next_latest_scan,
                        next_distance,
                        next_cos,
                        next_sin,
                        next_collision,
                        next_goal,
                        next_action,
                    )
                    reward = self.reward_function(
                        next_goal, next_collision, action, next_latest_scan
                    )
                    self.replay_buffer.add(
                        state, action, reward, next_terminal, next_state
                    )

        return self.replay_buffer

    def train(
        self,
        pretraining_iterations,
        replay_buffer,
        iterations,
        batch_size,
    ):
        """
        Run pretraining on the model using the replay buffer.

        Args:
            pretraining_iterations (int): Number of outer loop iterations for pretraining.
            replay_buffer (object): Buffer to sample training batches from.
            iterations (int): Number of training steps per pretraining iteration.
            batch_size (int): Batch size used during training.
        """
        print("Running Pretraining")
        for _ in tqdm(range(pretraining_iterations)):
            self.model.train(
                replay_buffer=replay_buffer,
                iterations=iterations,
                batch_size=batch_size,
            )
        print("Model Pretrained")

load_buffer()

Load samples from the specified files and populate the replay buffer.

Returns:

Name Type Description
object

The populated replay buffer.
Source code in robot_nav/utils.py 
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
def load_buffer(self):
    """
    Load samples from the specified files and populate the replay buffer.

    Returns:
        object: The populated replay buffer.
    """
    for file_name in self.file_names:
        print("Loading file: ", file_name)
        with open(file_name, "r") as file:
            samples = yaml.full_load(file)
            for i in tqdm(range(1, len(samples) - 1)):
                sample = samples[i]
                latest_scan = sample["latest_scan"]
                distance = sample["distance"]
                cos = sample["cos"]
                sin = sample["sin"]
                collision = sample["collision"]
                goal = sample["goal"]
                action = sample["action"]

                state, terminal = self.model.prepare_state(
                    latest_scan, distance, cos, sin, collision, goal, action
                )

                if terminal:
                    continue

                next_sample = samples[i + 1]
                next_latest_scan = next_sample["latest_scan"]
                next_distance = next_sample["distance"]
                next_cos = next_sample["cos"]
                next_sin = next_sample["sin"]
                next_collision = next_sample["collision"]
                next_goal = next_sample["goal"]
                next_action = next_sample["action"]
                next_state, next_terminal = self.model.prepare_state(
                    next_latest_scan,
                    next_distance,
                    next_cos,
                    next_sin,
                    next_collision,
                    next_goal,
                    next_action,
                )
                reward = self.reward_function(
                    next_goal, next_collision, action, next_latest_scan
                )
                self.replay_buffer.add(
                    state, action, reward, next_terminal, next_state
                )

    return self.replay_buffer

train(pretraining_iterations, replay_buffer, iterations, batch_size)

Run pretraining on the model using the replay buffer.

Parameters:

Name Type Description Default
pretraining_iterations int

Number of outer loop iterations for pretraining.

 required
replay_buffer object

Buffer to sample training batches from.

 required
iterations int

Number of training steps per pretraining iteration.

 required
batch_size int

Batch size used during training.

 required
Source code in robot_nav/utils.py 
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
def train(
    self,
    pretraining_iterations,
    replay_buffer,
    iterations,
    batch_size,
):
    """
    Run pretraining on the model using the replay buffer.

    Args:
        pretraining_iterations (int): Number of outer loop iterations for pretraining.
        replay_buffer (object): Buffer to sample training batches from.
        iterations (int): Number of training steps per pretraining iteration.
        batch_size (int): Batch size used during training.
    """
    print("Running Pretraining")
    for _ in tqdm(range(pretraining_iterations)):
        self.model.train(
            replay_buffer=replay_buffer,
            iterations=iterations,
            batch_size=batch_size,
        )
    print("Model Pretrained")

get_buffer(model, sim, load_saved_buffer, pretrain, pretraining_iterations, training_iterations, batch_size, buffer_size=50000, random_seed=666, file_names=['robot_nav/assets/data.yml'], history_len=10)

Get or construct the replay buffer depending on model type and training configuration.

Parameters:

Name Type Description Default
model object

The RL model, can be PPO, RCPG, or other.

 required
sim object

Simulation environment with a get_reward function.

 required
load_saved_buffer bool

Whether to load experiences from file.

 required
pretrain bool

Whether to run pretraining using the buffer.

 required
pretraining_iterations int

Number of outer loop iterations for pretraining.

 required
training_iterations int

Number of iterations in each training loop.

 required
batch_size int

Size of the training batch.

 required
buffer_size int

Maximum size of the buffer. Defaults to 50000.

 50000
random_seed int

Seed for reproducibility. Defaults to 666.

 666
file_names List[str]

List of YAML data file paths. Defaults to ["robot_nav/assets/data.yml"].

 ['robot_nav/assets/data.yml']
history_len int

Used for RCPG buffer configuration. Defaults to 10.

 10

Returns:

Name Type Description
object

The initialized and optionally pre-populated replay buffer.
Source code in robot_nav/utils.py 
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
def get_buffer(
    model,
    sim,
    load_saved_buffer,
    pretrain,
    pretraining_iterations,
    training_iterations,
    batch_size,
    buffer_size=50000,
    random_seed=666,
    file_names=["robot_nav/assets/data.yml"],
    history_len=10,
):
    """
    Get or construct the replay buffer depending on model type and training configuration.

    Args:
        model (object): The RL model, can be PPO, RCPG, or other.
        sim (object): Simulation environment with a `get_reward` function.
        load_saved_buffer (bool): Whether to load experiences from file.
        pretrain (bool): Whether to run pretraining using the buffer.
        pretraining_iterations (int): Number of outer loop iterations for pretraining.
        training_iterations (int): Number of iterations in each training loop.
        batch_size (int): Size of the training batch.
        buffer_size (int, optional): Maximum size of the buffer. Defaults to 50000.
        random_seed (int, optional): Seed for reproducibility. Defaults to 666.
        file_names (List[str], optional): List of YAML data file paths. Defaults to ["robot_nav/assets/data.yml"].
        history_len (int, optional): Used for RCPG buffer configuration. Defaults to 10.

    Returns:
        object: The initialized and optionally pre-populated replay buffer.
    """
    if isinstance(model, PPO):
        return model.buffer

    if isinstance(model, RCPG):
        replay_buffer = RolloutReplayBuffer(
            buffer_size=buffer_size, random_seed=random_seed, history_len=history_len
        )
    else:
        replay_buffer = ReplayBuffer(buffer_size=buffer_size, random_seed=random_seed)

    if pretrain:
        assert (
            load_saved_buffer
        ), "To pre-train model, load_saved_buffer must be set to True"

    if load_saved_buffer:
        pretraining = Pretraining(
            file_names=file_names,
            model=model,
            replay_buffer=replay_buffer,
            reward_function=sim.get_reward,
        )  # instantiate pre-trainind
        replay_buffer = (
            pretraining.load_buffer()
        )  # fill buffer with experiences from the data.yml file
        if pretrain:
            pretraining.train(
                pretraining_iterations=pretraining_iterations,
                replay_buffer=replay_buffer,
                iterations=training_iterations,
                batch_size=batch_size,
            )  # run pre-training

    return replay_buffer

get_max_bound(next_state, discount, max_ang_vel, max_lin_vel, time_step, distance_norm, goal_reward, reward, done, device)

Estimate the maximum possible return (upper bound) from the next state onward.

This is used in constrained RL or safe policy optimization where a conservative estimate of return is useful for policy updates.

Parameters:

Name Type Description Default
next_state Tensor

Tensor of next state observations.

 required
discount float

Discount factor for future rewards.

 required
max_ang_vel float

Maximum angular velocity of the agent.

 required
max_lin_vel float

Maximum linear velocity of the agent.

 required
time_step float

Duration of one time step.

 required
distance_norm float

Normalization factor for distance.

 required
goal_reward float

Reward received upon reaching the goal.

 required
reward Tensor

Immediate reward from the environment.

 required
done Tensor

Binary tensor indicating episode termination.

 required
device device

PyTorch device for computation.

 required

Returns:

Type Description

torch.Tensor: Maximum return bound for each sample in the batch.
Source code in robot_nav/utils.py 
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
def get_max_bound(
    next_state,
    discount,
    max_ang_vel,
    max_lin_vel,
    time_step,
    distance_norm,
    goal_reward,
    reward,
    done,
    device,
):
    """
    Estimate the maximum possible return (upper bound) from the next state onward.

    This is used in constrained RL or safe policy optimization where a conservative
    estimate of return is useful for policy updates.

    Args:
        next_state (torch.Tensor): Tensor of next state observations.
        discount (float): Discount factor for future rewards.
        max_ang_vel (float): Maximum angular velocity of the agent.
        max_lin_vel (float): Maximum linear velocity of the agent.
        time_step (float): Duration of one time step.
        distance_norm (float): Normalization factor for distance.
        goal_reward (float): Reward received upon reaching the goal.
        reward (torch.Tensor): Immediate reward from the environment.
        done (torch.Tensor): Binary tensor indicating episode termination.
        device (torch.device): PyTorch device for computation.

    Returns:
        torch.Tensor: Maximum return bound for each sample in the batch.
    """
    next_state = next_state.clone()  # Prevents in-place modifications
    reward = reward.clone()  # Ensures original reward is unchanged
    done = done.clone()
    cos = next_state[:, -4]
    sin = next_state[:, -3]
    theta = torch.atan2(sin, cos)

    # Compute turning steps
    turn_steps = theta.abs() / (max_ang_vel * time_step)
    full_turn_steps = torch.floor(turn_steps)
    turn_rew = -max_ang_vel * discount**full_turn_steps
    turn_rew[full_turn_steps == 0] = 0  # Handle zero case
    final_turn_rew = -(discount ** (full_turn_steps + 1)) * (
        turn_steps - full_turn_steps
    )
    full_turn_rew = turn_rew + final_turn_rew

    # Compute distance-based steps
    full_turn_steps += 1  # Account for the final turn step
    distances = (next_state[:, -5] * distance_norm) / (max_lin_vel * time_step)
    final_steps = torch.ceil(distances) + full_turn_steps
    inter_steps = torch.trunc(distances) + full_turn_steps

    final_rew = goal_reward * discount**final_steps

    # Compute intermediate rewards using a sum of discounted steps
    max_inter_steps = inter_steps.max().int().item()
    discount_exponents = discount ** torch.arange(1, max_inter_steps + 1, device=device)
    inter_rew = torch.stack(
        [
            (max_lin_vel * discount_exponents[int(start) + 1 : int(steps)]).sum()
            for start, steps in zip(full_turn_steps, inter_steps)
        ]
    )
    # Compute final max bound
    max_bound = reward + (1 - done) * (full_turn_rew + final_rew + inter_rew).view(
        -1, 1
    )
    return max_bound