Skip to content

Proximal Policy Gradient (PPO)

Overview

PPO is one of the most popular DRL algorithms. It runs reasonably fast by leveraging vector (parallel) environments and naturally works well with different action spaces, therefore supporting a variety of games. It also has good sample efficiency compared to algorithms such as DQN.

Original paper:

Reference resources:

All our PPO implementations below are augmented with the same code-level optimizations presented in openai/baselines's PPO. To achieve this, see how we matched the implementation details in our blog post The 37 Implementation Details of Proximal Policy Optimization.

Implemented Variants

Variants Implemented Description
ppo.py, docs For classic control tasks like CartPole-v1.
ppo_atari.py, docs For Atari games. It uses convolutional layers and common atari-based pre-processing techniques.
ppo_continuous_action.py, docs For continuous action space. Also implemented Mujoco-specific code-level optimizations
ppo_atari_lstm.py, docs For Atari games using LSTM without stacked frames.
ppo_atari_envpool.py, docs Uses the blazing fast Envpool Atari vectorized environment.
ppo_procgen.py, docs For the procgen environments
ppo_atari_multigpu.py, docs For Atari environments leveraging multi-GPUs
ppo_pettingzoo_ma_atari.py, docs For Pettingzoo's multi-agent Atari environments

Below are our single-file implementations of PPO:

ppo.py

The ppo.py has the following features:

  • Works with the Box observation space of low-level features
  • Works with the Discrete action space
  • Works with envs like CartPole-v1

Usage

poetry install
python cleanrl/ppo.py --help
python cleanrl/ppo.py --env-id CartPole-v1

Explanation of the logged metrics

Running python cleanrl/ppo.py will automatically record various metrics such as actor or value losses in Tensorboard. Below is the documentation for these metrics:

  • charts/episodic_return: episodic return of the game
  • charts/episodic_length: episodic length of the game
  • charts/SPS: number of steps per second
  • charts/learning_rate: the current learning rate
  • losses/value_loss: the mean value loss across all data points
  • losses/policy_loss: the mean policy loss across all data points
  • losses/entropy: the mean entropy value across all data points
  • losses/old_approx_kl: the approximate Kullback–Leibler divergence, measured by (-logratio).mean(), which corresponds to the k1 estimator in John Schulman’s blog post on approximating KL
  • losses/approx_kl: better alternative to olad_approx_kl measured by (logratio.exp() - 1) - logratio, which corresponds to the k3 estimator in approximating KL
  • losses/clipfrac: the fraction of the training data that triggered the clipped objective
  • losses/explained_variance: the explained variance for the value function

Implementation details

ppo.py is based on the "13 core implementation details" in The 37 Implementation Details of Proximal Policy Optimization, which are as follows:

  1. Vectorized architecture ( common/cmd_util.py#L22)
  2. Orthogonal Initialization of Weights and Constant Initialization of biases ( a2c/utils.py#L58))
  3. The Adam Optimizer's Epsilon Parameter ( ppo2/model.py#L100)
  4. Adam Learning Rate Annealing ( ppo2/ppo2.py#L133-L135)
  5. Generalized Advantage Estimation ( ppo2/runner.py#L56-L65)
  6. Mini-batch Updates ( ppo2/ppo2.py#L157-L166)
  7. Normalization of Advantages ( ppo2/model.py#L139)
  8. Clipped surrogate objective ( ppo2/model.py#L81-L86)
  9. Value Function Loss Clipping ( ppo2/model.py#L68-L75)
  10. Overall Loss and Entropy Bonus ( ppo2/model.py#L91)
  11. Global Gradient Clipping ( ppo2/model.py#L102-L108)
  12. Debug variables ( ppo2/model.py#L115-L116)
  13. Separate MLP networks for policy and value functions ( common/policies.py#L156-L160, baselines/common/models.py#L75-L103)

Experiment results

To run benchmark experiments, see benchmark/ppo.sh. Specifically, execute the following command:

Below are the average episodic returns for ppo.py. To ensure the quality of the implementation, we compared the results against openai/baselies' PPO.

Environment ppo.py openai/baselies' PPO (Huang et al., 2022)1
CartPole-v1 492.40 ± 13.05 497.54 ± 4.02
Acrobot-v1 -89.93 ± 6.34 -81.82 ± 5.58
MountainCar-v0 -200.00 ± 0.00 -200.00 ± 0.00

Learning curves:

Tracked experiments and game play videos:

Video tutorial

If you'd like to learn ppo.py in-depth, consider checking out the following video tutorial:

ppo_atari.py

The ppo_atari.py has the following features:

  • For Atari games. It uses convolutional layers and common atari-based pre-processing techniques.
  • Works with the Atari's pixel Box observation space of shape (210, 160, 3)
  • Works with the Discrete action space

Usage

poetry install -E atari
python cleanrl/ppo_atari.py --help
python cleanrl/ppo_atari.py --env-id BreakoutNoFrameskip-v4

Explanation of the logged metrics

See related docs for ppo.py.

Implementation details

ppo_atari.py is based on the "9 Atari implementation details" in The 37 Implementation Details of Proximal Policy Optimization, which are as follows:

  1. The Use of NoopResetEnv ( common/atari_wrappers.py#L12)
  2. The Use of MaxAndSkipEnv ( common/atari_wrappers.py#L97)
  3. The Use of EpisodicLifeEnv ( common/atari_wrappers.py#L61)
  4. The Use of FireResetEnv ( common/atari_wrappers.py#L41)
  5. The Use of WarpFrame (Image transformation) common/atari_wrappers.py#L134
  6. The Use of ClipRewardEnv ( common/atari_wrappers.py#L125)
  7. The Use of FrameStack ( common/atari_wrappers.py#L188)
  8. Shared Nature-CNN network for the policy and value functions ( common/policies.py#L157, common/models.py#L15-L26)
  9. Scaling the Images to Range [0, 1] ( common/models.py#L19)

Experiment results

To run benchmark experiments, see benchmark/ppo.sh. Specifically, execute the following command:

Below are the average episodic returns for ppo_atari.py. To ensure the quality of the implementation, we compared the results against openai/baselies' PPO.

Environment ppo_atari.py openai/baselies' PPO (Huang et al., 2022)1
BreakoutNoFrameskip-v4 416.31 ± 43.92 406.57 ± 31.554
PongNoFrameskip-v4 20.59 ± 0.35 20.512 ± 0.50
BeamRiderNoFrameskip-v4 2445.38 ± 528.91 2642.97 ± 670.37

Learning curves:

Tracked experiments and game play videos:

Video tutorial

If you'd like to learn ppo_atari.py in-depth, consider checking out the following video tutorial:

ppo_continuous_action.py

The ppo_continuous_action.py has the following features:

  • For continuous action space. Also implemented Mujoco-specific code-level optimizations
  • Works with the Box observation space of low-level features
  • Works with the Box (continuous) action space

Usage

poetry install -E atari
python cleanrl/ppo_continuous_action.py --help
python cleanrl/ppo_continuous_action.py --env-id Hopper-v2

Explanation of the logged metrics

See related docs for ppo.py.

Implementation details

ppo_continuous_action.py is based on the "9 details for continuous action domains (e.g. Mujoco)" in The 37 Implementation Details of Proximal Policy Optimization, which are as follows:

  1. Continuous actions via normal distributions ( common/distributions.py#L103-L104)
  2. State-independent log standard deviation ( common/distributions.py#L104)
  3. Independent action components ( common/distributions.py#L238-L246)
  4. Separate MLP networks for policy and value functions ( common/policies.py#L160, baselines/common/models.py#L75-L103
  5. Handling of action clipping to valid range and storage ( common/cmd_util.py#L99-L100)
  6. Normalization of Observation ( common/vec_env/vec_normalize.py#L4)
  7. Observation Clipping ( common/vec_env/vec_normalize.py#L39)
  8. Reward Scaling ( common/vec_env/vec_normalize.py#L28)
  9. Reward Clipping ( common/vec_env/vec_normalize.py#L32)

Experiment results

To run benchmark experiments, see benchmark/ppo.sh. Specifically, execute the following command:

Below are the average episodic returns for ppo_continuous_action.py. To ensure the quality of the implementation, we compared the results against openai/baselies' PPO.

Environment ppo_continuous_action.py openai/baselies' PPO (Huang et al., 2022)1
Hopper-v2 2231.12 ± 656.72 2518.95 ± 850.46
Walker2d-v2 3050.09 ± 1136.21 3208.08 ± 1264.37
HalfCheetah-v2 1822.82 ± 928.11 2152.26 ± 1159.84

Learning curves:

Tracked experiments and game play videos:

Video tutorial

If you'd like to learn ppo_continuous_action.py in-depth, consider checking out the following video tutorial:

ppo_atari_lstm.py

The ppo_atari_lstm.py has the following features:

  • For Atari games using LSTM without stacked frames. It uses convolutional layers and common atari-based pre-processing techniques.
  • Works with the Atari's pixel Box observation space of shape (210, 160, 3)
  • Works with the Discrete action space

Usage

poetry install -E atari
python cleanrl/ppo_atari_lstm.py --help
python cleanrl/ppo_atari_lstm.py --env-id BreakoutNoFrameskip-v4

Explanation of the logged metrics

See related docs for ppo.py.

Implementation details

ppo_atari_lstm.py is based on the "5 LSTM implementation details" in The 37 Implementation Details of Proximal Policy Optimization, which are as follows:

  1. Layer initialization for LSTM layers ( a2c/utils.py#L84-L86)
  2. Initialize the LSTM states to be zeros ( common/models.py#L179)
  3. Reset LSTM states at the end of the episode ( common/models.py#L141)
  4. Prepare sequential rollouts in mini-batches ( a2c/utils.py#L81)
  5. Reconstruct LSTM states during training ( a2c/utils.py#L81)

To help test out the memory, we remove the 4 stacked frames from the observation (i.e., using env = gym.wrappers.FrameStack(env, 1) instead of env = gym.wrappers.FrameStack(env, 4) like in ppo_atari.py )

Experiment results

To run benchmark experiments, see benchmark/ppo.sh. Specifically, execute the following command:

Below are the average episodic returns for ppo_atari_lstm.py. To ensure the quality of the implementation, we compared the results against openai/baselies' PPO.

Environment ppo_atari_lstm.py openai/baselies' PPO (Huang et al., 2022)1
BreakoutNoFrameskip-v4 128.92 ± 31.10 138.98 ± 50.76
PongNoFrameskip-v4 19.78 ± 1.58 19.79 ± 0.67
BeamRiderNoFrameskip-v4 1536.20 ± 612.21 1591.68 ± 372.95

Learning curves:

Tracked experiments and game play videos:

ppo_atari_envpool.py

The ppo_atari_envpool.py has the following features:

  • Uses the blazing fast Envpool vectorized environment.
  • For Atari games. It uses convolutional layers and common atari-based pre-processing techniques.
  • Works with the Atari's pixel Box observation space of shape (210, 160, 3)
  • Works with the Discrete action space
Warning

Note that ppo_atari_envpool.py does not work in Windows and MacOs . See envpool's built wheels here: https://pypi.org/project/envpool/#files

Usage

poetry install -E envpool
python cleanrl/ppo_atari_envpool.py --help
python cleanrl/ppo_atari_envpool.py --env-id Breakout-v5

Explanation of the logged metrics

See related docs for ppo.py.

Implementation details

ppo_atari_envpool.py uses a customized RecordEpisodeStatistics to work with envpool but has the same other implementation details as ppo_atari.py (see related docs).

Experiment results

To run benchmark experiments, see benchmark/ppo.sh. Specifically, execute the following command:

Below are the average episodic returns for ppo_atari_envpool.py. Notice it has the same sample efficiency as ppo_atari.py, but runs about 3x faster.

Environment ppo_atari_envpool.py (~80 mins) ppo_atari.py (~220 mins)
BreakoutNoFrameskip-v4 389.57 ± 29.62 416.31 ± 43.92
PongNoFrameskip-v4 20.55 ± 0.37 20.59 ± 0.35
BeamRiderNoFrameskip-v4 2039.83 ± 1146.62 2445.38 ± 528.91

Learning curves:

Tracked experiments and game play videos:

ppo_procgen.py

The ppo_procgen.py has the following features:

  • For the procgen environments
  • Uses IMPALA-style neural network
  • Works with the Discrete action space

Usage

poetry install -E procgen
python cleanrl/ppo_procgen.py --help
python cleanrl/ppo_procgen.py --env-id starpilot

Explanation of the logged metrics

See related docs for ppo.py.

Implementation details

ppo_procgen.py is based on the details in "Appendix" in The 37 Implementation Details of Proximal Policy Optimization, which are as follows:

  1. IMPALA-style Neural Network ( common/models.py#L28)
  2. Use the same gamma parameter in the NormalizeReward wrapper. Note that the original implementation from openai/train-procgen uses the default gamma=0.99 in the VecNormalize wrapper but gamma=0.999 as PPO's parameter. The mismatch between the gammas is technically incorrect. See #209

Experiment results

To run benchmark experiments, see benchmark/ppo.sh. Specifically, execute the following command:

We try to match the default setting in openai/train-procgen except that we use the easy distribution mode and total_timesteps=25e6 to save compute. Notice openai/train-procgen has the following settings:

  1. Learning rate annealing is turned off by default
  2. Reward scaling and reward clipping is used

Below are the average episodic returns for ppo_procgen.py. To ensure the quality of the implementation, we compared the results against openai/baselies' PPO.

Environment ppo_procgen.py openai/baselies' PPO (Huang et al., 2022)1
StarPilot (easy) 32.47 ± 11.21 33.97 ± 7.86
BossFight (easy) 9.63 ± 2.35 9.35 ± 2.04
BigFish (easy) 16.80 ± 9.49 20.06 ± 5.34
Info

Note that we have run the procgen experiments using the easy distribution for reducing the computational cost.

Learning curves:

Tracked experiments and game play videos:

ppo_atari_multigpu.py

The ppo_atari_multigpu.py leverages data parallelism to speed up training time at no cost of sample efficiency.

ppo_atari_multigpu.py has the following features:

  • Allows the users to use do training leveraging data parallelism
  • For playing Atari games. It uses convolutional layers and common atari-based pre-processing techniques.
  • Works with the Atari's pixel Box observation space of shape (210, 160, 3)
  • Works with the Discrete action space
Warning

Note that ppo_atari_multigpu.py does not work in Windows and MacOs . It will error out with NOTE: Redirects are currently not supported in Windows or MacOs. See pytorch/pytorch#20380

Usage

poetry install -E atari
python cleanrl/ppo_atari_multigpu.py --help

# `--nproc_per_node=2` specifies how many subprocesses we spawn for training with data parallelism
# note it is possible to run this with a *single GPU*: each process will simply share the same GPU
torchrun --standalone --nnodes=1 --nproc_per_node=2 cleanrl/ppo_atari_multigpu.py --env-id BreakoutNoFrameskip-v4

# by default we use the `gloo` backend, but you can use the `nccl` backend for better multi-GPU performance
torchrun --standalone --nnodes=1 --nproc_per_node=2 cleanrl/ppo_atari_multigpu.py --env-id BreakoutNoFrameskip-v4 --backend nccl

# it is possible to spawn more processes than the amount of GPUs you have via `--device-ids`
# e.g., the command below spawns two processes using GPU 0 and two processes using GPU 1
torchrun --standalone --nnodes=1 --nproc_per_node=2 cleanrl/ppo_atari_multigpu.py --env-id BreakoutNoFrameskip-v4 --device-ids 0 0 1 1

Explanation of the logged metrics

See related docs for ppo.py.

Implementation details

ppo_atari_multigpu.py is based on ppo_atari.py (see its related docs).

We use Pytorch's distributed API to implement the data parallelism paradigm. The basic idea is that the user can spawn \(N\) processes each holding a copy of the model, step the environments, and averages their gradients together for the backward pass. Here are a few note-worthy implementation details.

  1. Shard the environments: by default, ppo_atari_multigpu.py uses --num-envs=8. When calling torchrun --standalone --nnodes=1 --nproc_per_node=2 cleanrl/ppo_atari_multigpu.py --env-id BreakoutNoFrameskip-v4, it spawns \(N=2\) (by --nproc_per_node=2) subprocesses and shard the environments across these 2 subprocesses. In particular, each subprocess will have 8/2=4 environments. Implementation wise, we do args.num_envs = int(args.num_envs / world_size). Here world_size=2 refers to the size of the world, which means the group of subprocesses. We also need to adjust various variables as follows:
    • batch size: by default it is (num_envs * num_steps) = 8 * 128 = 1024 and we adjust it to (num_envs / world_size * num_steps) = (4 * 128) = 512.
    • minibatch size: by default it is (num_envs * num_steps) / num_minibatches = (8 * 128) / 4 = 256 and we adjust it to (num_envs / world_size * num_steps) / num_minibatches = (4 * 128) / 4 = 128.
    • number of updates: by default it is total_timesteps // batch_size = 10000000 // (8 * 128) = 9765 and we adjust it to total_timesteps // (batch_size * world_size) = 10000000 // (8 * 128 * 2) = 4882.
    • global step increment: by default it is num_envs and we adjust it to num_envs * world_size.
  2. Adjust seed per process: we need be very careful with seeding: we could have used the exact same seed for each subprocess. To ensure this does not happen, we do the following

    # CRUCIAL: note that we needed to pass a different seed for each data parallelism worker
    args.seed += local_rank
    random.seed(args.seed)
    np.random.seed(args.seed)
    torch.manual_seed(args.seed - local_rank)
    torch.backends.cudnn.deterministic = args.torch_deterministic
    
    # ...
    
    envs = gym.vector.SyncVectorEnv(
        [make_env(args.env_id, args.seed + i, i, args.capture_video, run_name) for i in range(args.num_envs)]
    )
    assert isinstance(envs.single_action_space, gym.spaces.Discrete), "only discrete action space is supported"
    
    agent = Agent(envs).to(device)
    torch.manual_seed(args.seed)
    optimizer = optim.Adam(agent.parameters(), lr=args.learning_rate, eps=1e-5)
    

    Notice that we adjust the seed with args.seed += local_rank (line 2), where local_rank is the index of the subprocesses. This ensures we seed packages and envs with uncorrealted seeds. However, we do need to use the same torch seed for all process to initialize same weights for the agent (line 5), after which we can use a different seed for torch (line 16). 1. Efficient gradient averaging: PyTorch recommends to average the gradient across the whole world via the following (see docs)

    for param in agent.parameters():
        dist.all_reduce(param.grad.data, op=dist.ReduceOp.SUM)
        param.grad.data /= world_size
    

    However, @cswinter introduces a more efficient gradient averaging scheme with proper batching (see entity-neural-network/incubator#220), which looks like:

    all_grads_list = []
    for param in agent.parameters():
        if param.grad is not None:
            all_grads_list.append(param.grad.view(-1))
    all_grads = torch.cat(all_grads_list)
    dist.all_reduce(all_grads, op=dist.ReduceOp.SUM)
    offset = 0
    for param in agent.parameters():
        if param.grad is not None:
            param.grad.data.copy_(
                all_grads[offset : offset + param.numel()].view_as(param.grad.data) / world_size
            )
            offset += param.numel()
    

    In our previous empirical testing (see vwxyzjn/cleanrl#162), we have found @cswinter's implementation to be faster, hence we adopt it in our implementation.

We can see how ppo_atari_multigpu.py can result in no loss of sample efficiency. In this example, the ppo_atari.py's minibatch size is 256 and the ppo_atari_multigpu.py's minibatch size is 128 with world size 2. Because we average gradient across the world, the gradient under ppo_atari_multigpu.py should be virtually the same as the gradient under ppo_atari.py.

Experiment results

To run benchmark experiments, see benchmark/ppo.sh. Specifically, execute the following command:

Below are the average episodic returns for ppo_atari_multigpu.py. To ensure no loss of sample efficiency, we compared the results against ppo_atari.py.

Environment ppo_atari_multigpu.py (in ~160 mins) ppo_atari.py (in ~215 mins)
BreakoutNoFrameskip-v4 429.06 ± 52.09 416.31 ± 43.92
PongNoFrameskip-v4 20.40 ± 0.46 20.59 ± 0.35
BeamRiderNoFrameskip-v4 2454.54 ± 740.49 2445.38 ± 528.91

Learning curves:

Under the same hardware, we see that ppo_atari_multigpu.py is about 30% faster than ppo_atari.py with no loss of sample efficiency.

Info

Although ppo_atari_multigpu.py is 30% faster than ppo_atari.py, ppo_atari_multigpu.py is still slower than ppo_atari_envpool.py, as shown below. This comparison really highlights the different kinds of optimization possible.

The purpose of ppo_atari_multigpu.py is not (yet) to achieve the fastest PPO + Atari example. Rather, its purpose is to rigorously validate data paralleism does provide performance benefits. We could do something like ppo_atari_multigpu_envpool.py to possibly obtain the fastest PPO + Atari possible, but that is for another day. Note we may need numba to pin the threads envpool is using in each subprocess to avoid threads fighting each other and lowering the throughput.

Tracked experiments and game play videos:

ppo_pettingzoo_ma_atari.py

ppo_pettingzoo_ma_atari.py trains an agent to learn playing Atari games via selfplay. The selfplay environment is implemented as a vectorized environment from PettingZoo.ml. The basic idea is to create vectorized environment \(E\) with num_envs = N, where \(N\) is the number of players in the game. Say \(N = 2\), then the 0-th sub environment of \(E\) will return the observation for player 0 and 1-th sub environment will return the observation of player 1. Then the two environments takes a batch of 2 actions and execute them for player 0 and player 1, respectively. See "Vectorized architecture" in The 37 Implementation Details of Proximal Policy Optimization for more detail.

ppo_pettingzoo_ma_atari.py has the following features:

  • For playing the pettingzoo's multi-agent Atari game.
  • Works with the pixel-based observation space
  • Works with the Box action space
Warning

Note that ppo_pettingzoo_ma_atari.py does not work in Windows . See https://pypi.org/project/multi-agent-ale-py/#files

Usage

poetry install -E "pettingzoo atari"
poetry run AutoROM --accept-license
python cleanrl/ppo_pettingzoo_ma_atari.py --help
python cleanrl/ppo_pettingzoo_ma_atari.py --env-id pong_v3
python cleanrl/ppo_pettingzoo_ma_atari.py --env-id surround_v2

See https://www.pettingzoo.ml/atari for a full-list of supported environments such as basketball_pong_v3. Notice pettingzoo sometimes introduces breaking changes, so make sure to install the pinned dependencies via poetry.

Explanation of the logged metrics

Additionally, it logs the following metrics

  • charts/episodic_return-player0: episodic return of the game for player 0
  • charts/episodic_return-player1: episodic return of the game for player 1
  • charts/episodic_length-player0: episodic length of the game for player 0
  • charts/episodic_length-player1: episodic length of the game for player 1

See other logged metrics in the related docs for ppo.py.

Implementation details

ppo_pettingzoo_ma_atari.py is based on ppo_atari.py (see its related docs).

ppo_pettingzoo_ma_atari.py additionally has the following implementation details:

  1. supersuit wrappers: uses preprocessing wrappers from supersuit instead of from stable_baselines3, which looks like the following. In particular note that the supersuit does not offer a wrapper similar to NoopResetEnv, and that it uses the agent_indicator_v0 to add two channels indicating the which player the agent controls.

    -env = gym.make(env_id)
    -env = NoopResetEnv(env, noop_max=30)
    -env = MaxAndSkipEnv(env, skip=4)
    -env = EpisodicLifeEnv(env)
    -if "FIRE" in env.unwrapped.get_action_meanings():
    -    env = FireResetEnv(env)
    -env = ClipRewardEnv(env)
    -env = gym.wrappers.ResizeObservation(env, (84, 84))
    -env = gym.wrappers.GrayScaleObservation(env)
    -env = gym.wrappers.FrameStack(env, 4)
    +env = importlib.import_module(f"pettingzoo.atari.{args.env_id}").parallel_env()
    +env = ss.max_observation_v0(env, 2)
    +env = ss.frame_skip_v0(env, 4)
    +env = ss.clip_reward_v0(env, lower_bound=-1, upper_bound=1)
    +env = ss.color_reduction_v0(env, mode="B")
    +env = ss.resize_v1(env, x_size=84, y_size=84)
    +env = ss.frame_stack_v1(env, 4)
    +env = ss.agent_indicator_v0(env, type_only=False)
    +env = ss.pettingzoo_env_to_vec_env_v1(env)
    +envs = ss.concat_vec_envs_v1(env, args.num_envs // 2, num_cpus=0, base_class="gym")
    
    1. A more detailed note on the agent_indicator_v0 wrapper: let's dig deeper into how agent_indicator_v0 works. We do print(envs.reset(), envs.reset().shape)
    [  0.,   0.,   0., 236.,   1,   0.]],
    
    [[  0.,   0.,   0., 236.,   0.,   1.],
    [  0.,   0.,   0., 236.,   0.,   1.],
    [  0.,   0.,   0., 236.,   0.,   1.],
    ...,
    [  0.,   0.,   0., 236.,   0.,   1.],
    [  0.,   0.,   0., 236.,   0.,   1.],
    [  0.,   0.,   0., 236.,   0.,   1.]]]]) torch.Size([16, 84, 84, 6])
    

    So the agent_indicator_v0 adds the last two columns, where [ 0., 0., 0., 236., 1, 0.]] means this observation is for player 0, and [ 0., 0., 0., 236., 0., 1.] is for player 1. Notice the observation still has the range of \([0, 255]\) but the agent indicator channel has the range of \([0,1]\), so we need to be careful when dividing the observation by 255. In particular, we would only divide the first four channels by 255 and leave the agent indicator channels untouched as follows:

    def get_action_and_value(self, x, action=None):
        x = x.clone()
        x[:, :, :, [0, 1, 2, 3]] /= 255.0
        hidden = self.network(x.permute((0, 3, 1, 2)))
    

Experiment results

To run benchmark experiments, see benchmark/ppo.sh. Specifically, execute the following command:

Info

Note that evaluation is usually tricker in in selfplay environments. The usual episodic return is not a good indicator of the agent's performance in zero-sum games because the episodic return converges to zero. To evaluate the agent's ability, an intuitive approach is to take a look at the videos of the agents playing the game (included below), visually inspect the agent's behavior. The best scheme, however, is rating systems like Trueskill or ELO scores. However, they are more difficult to implement and are outside the scode of ppo_pettingzoo_ma_atari.py.

For simplicity, we measure the episodic length instead, which in a sense measures how many "back and forth" the agent can create. In other words, the longer the agent can play the game, the better the agent can play. Empirically, we have found episodic length to be a good indicator of the agent's skill, especially in pong_v3 and surround_v2. However, it is not the case for tennis_v3 and we'd need to visually inspect the agents' game play videos.

Below are the average episodic length for ppo_pettingzoo_ma_atari.py. To ensure no loss of sample efficiency, we compared the results against ppo_atari.py.

Environment ppo_pettingzoo_ma_atari.py
pong_v3 4153.60 ± 190.80
surround_v2 3055.33 ± 223.68
tennis_v3 14538.02 ± 7005.54

Learning curves:

Tracked experiments and game play videos:

ppo_continuous_action_isaacgym.py

Experimental

The ppo_continuous_action_isaacgym.py has the following features:

  • Works with IsaacGymEnvs.
  • Works with the Box observation space of low-level features
  • Works with the Box (continuous) action space

IsaacGymEnvs is a hardware-accelerated (or GPU-accelerated) robotics simulation environment based on torch, which allows us to run thousands of simulation environments at the same time, empowering RL agents to learn many MuJoCo-style robotics tasks in minutes instead of hours. When creating an environment with IsaacGymEnvs via isaacgymenvs.make("Ant"), it creates a vectorized environment which produces GPU tensors as observations and take GPU tensors as actions to execute.

Info

Note that Isaac Gym is the underlying core physics engine, and IssacGymEnvs is a collection of environments built on Isaac Gym.

Info

ppo_continuous_action_isaacgym.py works with most environments in IsaacGymEnvs but it does not work with the following environments yet:

  • AnymalTerrain
  • FrankaCabinet
  • ShadowHandOpenAI_FF
  • ShadowHandOpenAI_LSTM
  • Trifinger
  • Ingenuity Quadcopter

🔥 we need contributors to work on supporting and tuning our PPO implementation in these envs. If you are interested, please read our contribution guide and reach out!

Usage

The installation of isaacgym requires a bit of work since it's not a standard Python package.

Please go to https://developer.nvidia.com/isaac-gym to download and install the latest version of Issac Gym which should look like IsaacGym_Preview_4_Package.tar.gz. Put this IsaacGym_Preview_4_Package.tar.gz into the ~/Downloads/ folder. Make sure your python version is either 3.7, or 3.8 (3.9 not supported yet).

# extract and move the content in `python` folder in the IsaacGym_Preview_4_Package.tar.gz
# into the `cleanrl/ppo_continuous_action_isaacgym/isaacgym/` folder
cp ~/Downloads/IsaacGym_Preview_4_Package.tar.gz IsaacGym_Preview_4_Package.tar.gz 
stat IsaacGym_Preview_4_Package.tar.gz
mkdir temp_isaacgym
tar -xf IsaacGym_Preview_4_Package.tar.gz -C temp_isaacgym
mv temp_isaacgym/isaacgym/python/* cleanrl/ppo_continuous_action_isaacgym/isaacgym
rm -rf temp_isaacgym

# if your global python version is not either 3.7 nor 3.8, you need to tell poetry specifically to use a 3.7 or 3.8 python
# e.g., `poetry env use /home/costa/.pyenv/versions/3.7.8/bin/python`
poetry install -E isaacgym
# if you are using NVIDIA's 30xx GPU, you need to specifically install cuda 11.3 wheels
# `poetry run pip install torch --upgrade --extra-index-url https://download.pytorch.org/whl/cu113`
poetry run python cleanrl/ppo_continuous_action_isaacgym/ppo_continuous_action_isaacgym.py --help
poetry run python cleanrl/ppo_continuous_action_isaacgym/ppo_continuous_action_isaacgym.py --env-id Ant
Warning

If you encounter the following installation error

Python.h: No such file or directory
#include <Python.h>

or

libpython3.8.so.1.0: cannot open shared object file: No such file or directory

It usually means your python distribution does not include the shared library files. If you are ubuntu, you can install the following packages:

sudo apt-get install libpython3.8-dev # or sudo apt-get install libpython3.7-dev

If you are using pyenv, you may try the following:

env PYTHON_CONFIGURE_OPTS="--enable-shared" pyenv install 3.7.8

Explanation of the logged metrics

See related docs for ppo.py.

Additionally, charts/consecutive_successes means the number of consecutive episodes that the agent has successfully manipulating the rubix cube to the desired state.

Implementation details

ppo_continuous_action_isaacgym.py is based on ppo_continuous_action.py (see related docs), with a few modifications:

  1. Different set of hyperparameters: ppo_continuous_action_isaacgym.py uses hyperparameters primarily derived from rl-games' configuration (see example). The basic spirit is to run more total_timesteps, with larger num_envs and smaller num_steps.
arguments ppo_continuous_action.py ppo_continuous_action_isaacgym.py ppo_continuous_action_isaacgym.py (for ShadowHand and AllegroHand)
--total-timesteps 1000000 30000000 600000000
--learning-rate 3e-4 0.0026 0.0026
--num-envs 1 4096 8192
--num-steps 2048 16 8
--anneal-lr True False False
--num-minibatches 32 2 4
--update-epochs 10 4 5
--clip-vloss True False False
--vf-coef 0.5 2 2
--max-grad-norm 0.5 1 1
--reward-scaler N/A 1 0.01
  1. Slightly larger NN: ppo_continuous_action.py uses the following NN:
    self.critic = nn.Sequential(
        layer_init(nn.Linear(np.array(envs.single_observation_space.shape).prod(), 64)),
        nn.Tanh(),
        layer_init(nn.Linear(64, 64)),
        nn.Tanh(),
        layer_init(nn.Linear(64, 1), std=1.0),
    )
    self.actor_mean = nn.Sequential(
        layer_init(nn.Linear(np.array(envs.single_observation_space.shape).prod(), 64)),
        nn.Tanh(),
        layer_init(nn.Linear(64, 64)),
        nn.Tanh(),
        layer_init(nn.Linear(64, np.prod(envs.single_action_space.shape)), std=0.01),
    )
    
    while ppo_continuous_action_isaacgym.py uses the following NN:
    self.critic = nn.Sequential(
        layer_init(nn.Linear(np.array(envs.single_observation_space.shape).prod(), 256)),
        nn.Tanh(),
        layer_init(nn.Linear(256, 256)),
        nn.Tanh(),
        layer_init(nn.Linear(256, 1), std=1.0),
    )
    self.actor_mean = nn.Sequential(
        layer_init(nn.Linear(np.array(envs.single_observation_space.shape).prod(), 256)),
        nn.Tanh(),
        layer_init(nn.Linear(256, 256)),
        nn.Tanh(),
        layer_init(nn.Linear(256, np.prod(envs.single_action_space.shape)), std=0.01),
    )
    
  2. No normalization and clipping: ppo_continuous_action_isaacgym.py does not do observation and reward normalization and clipping for simplicity. It does however optionally offer an option to scale the rewards via --reward-scaler x, which multiplies all the rewards obtained by x as an example.
  3. Remove all CPU-related code: ppo_continuous_action_isaacgym.py needs to remove all CPU-related code (e.g. action.cpu().numpy()). This is because almost everything in IsaacGymEnvs happens in GPU. To do this, the major modifications include the following:
  4. Create a custom RecordEpisodeStatisticsTorch wrapper that records statstics using GPU tensors instead of numpy arrays.
  5. Avoid transferring the tensors to CPU. The related code in ppo_continuous_action.py looks like
    next_obs, reward, done, info = envs.step(action.cpu().numpy())
    rewards[step] = torch.tensor(reward).to(device).view(-1)
    next_obs, next_done = torch.Tensor(next_obs).to(device), torch.Tensor(done).to(device)
    
    and the related code in ppo_continuous_action_isaacgym.py looks like
    next_obs, rewards[step], next_done, info = envs.step(action)
    

Experiment results

To run benchmark experiments, see benchmark/ppo.sh. Specifically, execute the following command:

Below are the average episodic returns for ppo_continuous_action_isaacgym.py. To ensure the quality of the implementation, we compared the results against Denys88/rl_games' PPO and present the training time (units being s (seconds), m (minutes)). The hardware used is a NVIDIA RTX A6000 in a 24 core machine.

Environment (training time) ppo_continuous_action_isaacgym.py Denys88/rl_games
Cartpole (40s) 413.66 ± 120.93 417.49 (30s)
Ant (240s) 3953.30 ± 667.086 5873.05
Humanoid (350s) 2987.95 ± 257.60 6254.73
Anymal (317s) 29.34 ± 17.80 62.76
BallBalance (160s) 161.92 ± 89.20 319.76
AllegroHand (200m) 762.93 ± 427.92 3479.85
ShadowHand (130m) 427.16 ± 161.79 5713.74

Learning curves:

Info

Note ppo_continuous_action_isaacgym.py's performance seems poor compared to Denys88/rl_games' PPO. This is likely due to a few reasons.

  1. Denys88/rl_games' PPO uses different sets of tuned hyperparameters and neural network architecture configuration for different tasks, whereas ppo_continuous_action_isaacgym.py only uses one neural network architecture and 2 set of hyperparameters (ignoring --total-timesteps).
  2. ppo_continuous_action_isaacgym.py does not use observation normalization (because in my preliminary testing for some reasons it did not help).

While it should be possible to obtain higher scores with more tuning, the purpose of ppo_continuous_action_isaacgym.py is to hit a balance between simplicity and performance. I think ppo_continuous_action_isaacgym.py has relatively good performance with a concise codebase, which should be easy to modify and extend for practitioners.

Tracked experiments and game play videos:

Old Learning curves w/ Isaac Gym Preview 3 (no longer available in Nvidia's website for download):

Info

Note the AllegroHand and ShadowHand experiments used the following command ppo_continuous_action_isaacgym.py --track --capture-video --num-envs 16384 --num-steps 8 --update-epochs 5 --reward-scaler 0.01 --total-timesteps 600000000 --record-video-step-frequency 3660. Costa: I was able to run this during my internship at NVIDIA, but in my home setup, the computer has less GPU memory which makes it hard to replicate the results w/ --num-envs 16384.


  1. Huang, Shengyi; Dossa, Rousslan Fernand Julien; Raffin, Antonin; Kanervisto, Anssi; Wang, Weixun (2022). The 37 Implementation Details of Proximal Policy Optimization. ICLR 2022 Blog Track https://iclr-blog-track.github.io/2022/03/25/ppo-implementation-details/