Train Biped Robot to Walk Using Reinforcement Learning Agents

Fix Random Seed Generator to Improve Reproducibility

The example code may involve computation of random numbers at various stages such as initialization of the agent, creation of the actor and critic, resetting the environment during simulations, generating observations (for stochastic environments), generating exploration actions, and sampling min-batches of experiences for learning. Fixing the random number stream preserves the sequence of the random numbers every time you run the code and improves reproducibility of results. You will fix the random number stream at various locations in the example.

Fix the random number stream with the seed 0 and random number algorithm Mersenne Twister. For more information on random number generation see rng.

previousRngState = rng(0,"twister")

previousRngState = struct with fields:
     Type: 'twister'
     Seed: 0
    State: [625x1 uint32]

The output previousRngState is a structure that contains information about the previous state of the stream. You will restore the state at the end of the example.

Biped Robot Model

The reinforcement learning environment for this example is a biped robot. The training goal is to make the robot walk in a straight line using minimal control effort.

Load the parameters of the model into the MATLAB® workspace using the robotParametersRL Script provided in the example folder.

robotParametersRL

Open the Simulink® model.

mdl = "rlWalkingBipedRobot";
open_system(mdl)

The robot is modeled using Simscape™ Multibody™.

For this model:

The environment provides the following 29 observations to the agent.

The episode terminates if either of the following conditions occur.

The following reward function ${\mathit{r}}_{\mathit{t}}$, which is provided at every time step is inspired by [2].

Here:

This reward function encourages the agent to move forward by providing a positive reward for positive forward velocity. It also encourages the agent to avoid episode termination by providing a constant reward ($25\frac{\mathrm{Ts}}{\mathrm{Tf}}$) at every time step. The other terms in the reward function are penalties for substantial changes in lateral and vertical translations, and for the use of excess control effort.

Create Environment Object

Create the observation specification.

numObs = 29;
obsInfo = rlNumericSpec([numObs 1]);
obsInfo.Name = "observations";

Create the action specification.

numAct = 6;
actInfo = rlNumericSpec([numAct 1],LowerLimit=-1,UpperLimit=1);
actInfo.Name = "foot_torque";

Create the environment interface for the walking robot model.

blk = mdl + "/RL Agent";
env = rlSimulinkEnv(mdl,blk,obsInfo,actInfo);
env.ResetFcn = @(in) walkerResetFcn(in, ...
    upper_leg_length/100, ...
    lower_leg_length/100, ...
    h/100);

Select and Create Agent for Training

When you create the agent, the initial parameters of the actor and critic networks are initialized with random values. Fix the random number stream so that the agent is always initialized with the same parameter values.

rng(0,"twister");

This example provides the option to train the robot either using either a DDPG or TD3 agent. To simulate the robot with the agent of your choice, set the AgentSelection flag accordingly.

AgentSelection = "DDPG";
switch AgentSelection
    case "DDPG"
        agent = createDDPGAgent(numObs,obsInfo,numAct,actInfo,Ts);
    case "TD3"
        agent = createTD3Agent(numObs,obsInfo,numAct,actInfo,Ts);
    otherwise
        disp("Assign AgentSelection to DDPG or TD3")
end

The createDDPGAgent and createTD3Agent helper functions (provided in the example folder) perform the following actions.

DDPG Agent

DDPG agents use a parametrized deterministic policy over continuous action spaces, which is learned by a continuous deterministic actor, and a parametrized Q-value function approximator to estimate the value of the policy. Use neural networks to model both the policy and the Q-value function. The actor and critic networks for this example are inspired by [1].

For details on how the DDPG agent is created, see the createDDPGAgent helper function. For information on configuring DDPG agent options, see rlDDPGAgentOptions (Reinforcement Learning Toolbox).

For more information on creating a deep neural network value function representation, see Create Policies and Value Functions (Reinforcement Learning Toolbox). For an example that creates neural networks for DDPG agents, see Compare DDPG Agent to LQR Controller (Reinforcement Learning Toolbox).

TD3 Agent

The critic of a DDPG agent can overestimate the Q value. Since the agent uses the Q value to update its policy (actor), the resultant policy can be suboptimal and can accumulate training errors that can lead to divergent behavior. The TD3 algorithm is an extension of DDPG with improvements that make it more robust by preventing overestimation of Q values [3].

The structure of the actor and critic networks used for this agent are the same as the ones used for DDPG agent. For details on the creation of the TD3 agent, see the createTD3Agent helper function. For information on configuring TD3 agent options, see rlTD3AgentOptions (Reinforcement Learning Toolbox).

Get the actor and critic neural networks from the agent.

actor = getActor(agent);
critic = getCritic(agent);
actorNet = getModel(actor);
criticNet = getModel(critic(1));

View a summary and plot of the actor neural network.

summary(actorNet);

   Initialized: true

   Number of learnables: 134.1k

   Inputs:
      1   'observation'   29 features

plot(actorNet);

View a summary and plot of the actor neural network.

summary(criticNet);

   Initialized: true

   Number of learnables: 134.7k

   Inputs:
      1   'observation'   29 features
      2   'action'        6 features

plot(criticNet);

Specify Training Options and Train Agent

For this example, the training options for the DDPG and TD3 agents are the same.

For more information and additional options, see rlTrainingOptions (Reinforcement Learning Toolbox).

maxEpisodes = 3000;
maxSteps = floor(Tf/Ts);
trainOpts = rlTrainingOptions(...
    MaxEpisodes=maxEpisodes,...
    MaxStepsPerEpisode=maxSteps,...
    ScoreAveragingWindowLength=250,...
    StopTrainingCriteria="none",...
    SimulationStorageType="file",...
    SaveSimulationDirectory=AgentSelection+"Sims");

To train the agent in parallel, specify the following training options. Training in parallel requires Parallel Computing Toolbox™. If you do not have Parallel Computing Toolbox software installed, set UseParallel to false.

trainOpts.UseParallel = true;
trainOpts.ParallelizationOptions.Mode = "async";

In parallel training, workers simulate the agent's policy with the environment and store experiences in the replay memory. When workers operate asynchronously the order of stored experiences may not be deterministic and can ultimately make the training results different. To maximize the reproducibility likelihood:

Fix the random stream for reproducibility.

rng(0,"twister");

Train the agent using the train (Reinforcement Learning Toolbox) function. This process is computationally intensive and takes several hours to complete for each agent. To save time while running this example, load a pretrained agent by setting doTraining to false. To train the agent yourself, set doTraining to true. Due to randomness in the parallel training, you can expect different training results from the plots that follow. The pretrained agents were trained in parallel using four workers.

doTraining = false;
if doTraining    
    % Train the agent.
    trainingStats = train(agent,env,trainOpts);
else
    % Load a pretrained agent for the selected agent type.
    if strcmp(AgentSelection,"DDPG")
       load(fullfile("DDPGAgent","run1.mat"),"agent")
    else
       load(fullfile("TD3Agent","run1.mat"),"agent")
    end  
end

Simulate Trained Agents

Fix the random stream for reproducibility.

rng(0,"twister");

To validate the performance of the trained agent, simulate it within the biped robot environment. For more information on agent simulation, see rlSimulationOptions (Reinforcement Learning Toolbox) and sim (Reinforcement Learning Toolbox).

simOptions = rlSimulationOptions(MaxSteps=maxSteps);
experience = sim(env,agent,simOptions);

Compare Agent Performance

For the following agent comparison, each agent was trained five times using a different random seed each time. Due to the random exploration noise and the randomness in the parallel training, the learning curve for each run is different. Since the training of agents for multiple runs can take several days to complete, this comparison uses pretrained agents.

For the DDPG and TD3 agents, plot the average and standard deviation of the episode reward (top plot) and the episode Q0 value (bottom plot). The episode Q0 value is the critic estimate of the discounted long-term reward at the start of each episode given the initial observation of the environment. For a well-designed critic, the episode Q0 value approaches the true discounted long-term reward.

Fix the random stream for reproducibility.

rng(0,"twister");

Compare the performance of the pretrained agents using data stored in the DDPGAgent and TD3Agent folders. To plot the performance metrics, specify the folder names as input arguments to the comparePerformance function. The comparison data and the function are found in the example directory.

comparePerformance("DDPGAgent","TD3Agent")

Based on the Learning curve comparison plot:

Based on the Episode Q0 comparison plot:

For another example on how to train a humanoid robot to walk using a DDPG agent, see Train Humanoid Walker (Simscape Multibody). For an example on how to train a quadruped robot to walk using a DDPG agent, see Quadruped Robot Locomotion Using DDPG Agent (Reinforcement Learning Toolbox).

Restore the random number stream using the information stored in previousRngState.

rng(previousRngState);

References

[1] Lillicrap, Timothy P., Jonathan J. Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. "Continuous Control with Deep Reinforcement Learning." Preprint, submitted July 5, 2019. https://arxiv.org/abs/1509.02971.

[2] Heess, Nicolas, Dhruva TB, Srinivasan Sriram, Jay Lemmon, Josh Merel, Greg Wayne, Yuval Tassa, et al. "Emergence of Locomotion Behaviours in Rich Environments." Preprint, submitted July 10, 2017. https://arxiv.org/abs/1707.02286.

[3] Fujimoto, Scott, Herke van Hoof, and David Meger. "Addressing Function Approximation Error in Actor-Critic Methods." Preprint, submitted October 22, 2018. https://arxiv.org/abs/1802.09477.