simulation/modules/IsaacLab/scripts/benchmarks/benchmark_rlgames.py

# Copyright (c) 2022-2026, The Isaac Lab Project Developers (https://github.com/isaac-sim/IsaacLab/blob/main/CONTRIBUTORS.md).
# All rights reserved.
#
# SPDX-License-Identifier: BSD-3-Clause

"""Script to benchmark RL agent with RL-Games."""

"""Launch Isaac Sim Simulator first."""

import argparse
import contextlib
import os
import sys
import time

from isaaclab.app import AppLauncher

from isaaclab_tasks.utils import setup_preset_cli

from scripts.benchmarks.early_stop import (
    RlGamesEarlyStopObserver,
    add_success_cli_args,
    build_success_kwargs,
    get_success_tracker,
)

# add argparse arguments
parser = argparse.ArgumentParser(description="Train an RL agent with RL-Games.")
parser.add_argument("--video", action="store_true", default=False, help="Record videos during training.")
parser.add_argument("--video_length", type=int, default=200, help="Length of the recorded video (in steps).")
parser.add_argument("--video_interval", type=int, default=2000, help="Interval between video recordings (in steps).")
parser.add_argument("--num_envs", type=int, default=None, help="Number of environments to simulate.")
parser.add_argument("--task", type=str, default=None, help="Name of the task.")
parser.add_argument("--seed", type=int, default=None, help="Seed used for the environment")
parser.add_argument(
    "--distributed", action="store_true", default=False, help="Run training with multiple GPUs or nodes."
)
parser.add_argument("--max_iterations", type=int, default=10, help="RL Policy training iterations.")
parser.add_argument(
    "--benchmark_backend",
    type=str,
    default="omniperf",
    choices=[
        "json",
        "osmo",
        "omniperf",
        "summary",
        "LocalLogMetrics",
        "JSONFileMetrics",
        "OsmoKPIFile",
        "OmniPerfKPIFile",
    ],
    help="Benchmarking backend options, defaults omniperf",
)
parser.add_argument("--output_path", type=str, default=".", help="Path to output benchmark results.")
parser.add_argument(
    "--reward_threshold", type=float, default=None, help="Reward threshold for convergence (overrides config)."
)
parser.add_argument(
    "--check_convergence", action="store_true", help="Check reward convergence using thresholds from configs.yaml."
)
parser.add_argument(
    "--convergence_config", type=str, default="full", help="Config mode for convergence thresholds (default: full)."
)
add_success_cli_args(parser)

# append AppLauncher cli args
AppLauncher.add_app_launcher_args(parser)
args_cli, hydra_args = setup_preset_cli(parser)
sys.argv = [sys.argv[0]] + hydra_args
if args_cli.video:
    args_cli.enable_cameras = True

imports_time_begin = time.perf_counter_ns()

import math
import random
from datetime import datetime

import gymnasium as gym
import torch
from rl_games.common import env_configurations, vecenv
from rl_games.common.algo_observer import IsaacAlgoObserver
from rl_games.torch_runner import Runner

from isaaclab.envs import DirectMARLEnvCfg, DirectRLEnvCfg, ManagerBasedRLEnvCfg
from isaaclab.utils.dict import print_dict
from isaaclab.utils.io import dump_yaml

from isaaclab_rl.rl_games import RlGamesGpuEnv, RlGamesVecEnvWrapper

import isaaclab_tasks  # noqa: F401

# PLACEHOLDER: Extension template (do not remove this comment)
with contextlib.suppress(ImportError):
    import isaaclab_tasks_experimental  # noqa: F401
from isaaclab_tasks.utils import launch_simulation, resolve_task_config

imports_time_end = time.perf_counter_ns()

sys.path.append(os.path.join(os.path.dirname(os.path.abspath(__file__)), "../.."))

from isaaclab.test.benchmark import BaseIsaacLabBenchmark, BenchmarkMonitor
from isaaclab.utils.timer import Timer

from scripts.benchmarks.utils import (
    get_backend_type,
    get_preset_string,
    log_app_start_time,
    log_python_imports_time,
    log_rl_training_metrics,
    log_runtime_step_times,
    log_scene_creation_time,
    log_simulation_start_time,
    log_success,
    log_task_start_time,
    log_total_start_time,
    parse_tf_logs,
)

torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
torch.backends.cudnn.deterministic = False
torch.backends.cudnn.benchmark = False


# Create the benchmark
backend_type = get_backend_type(args_cli.benchmark_backend)
benchmark = BaseIsaacLabBenchmark(
    benchmark_name="benchmark_rlgames_train",
    backend_type=backend_type,
    output_path=args_cli.output_path,
    use_recorders=True,
    frametime_recorders=backend_type in ("summary", "omniperf"),
    output_prefix=f"benchmark_rlgames_train_{args_cli.task}",
    workflow_metadata={
        "metadata": [
            {"name": "task", "data": args_cli.task},
            {"name": "seed", "data": args_cli.seed},
            {"name": "num_envs", "data": args_cli.num_envs},
            {"name": "max_iterations", "data": args_cli.max_iterations},
            {"name": "presets", "data": get_preset_string(hydra_args)},
        ]
    },
)


def main(
    env_cfg: ManagerBasedRLEnvCfg | DirectRLEnvCfg | DirectMARLEnvCfg,
    agent_cfg: dict,
    app_start_time_begin: int,
    app_start_time_end: int,
):
    """Train with RL-Games agent."""

    # override configurations with non-hydra CLI arguments
    env_cfg.scene.num_envs = args_cli.num_envs if args_cli.num_envs is not None else env_cfg.scene.num_envs
    # For distributed training, launch_simulation() already resolved the
    # correct per-rank device; only apply a CLI --device override for
    # non-distributed runs (the default "cuda:0" would clobber the
    # per-rank device otherwise).
    if not args_cli.distributed:
        env_cfg.sim.device = args_cli.device if args_cli.device is not None else env_cfg.sim.device
    # check for invalid combination of CPU device with distributed training
    if args_cli.distributed and args_cli.device is not None and "cpu" in args_cli.device:
        raise ValueError(
            "Distributed training is not supported when using CPU device. "
            "Please use GPU device (e.g., --device cuda) for distributed training."
        )

    # update agent device to match simulation device (skip for distributed —
    # the per-rank device is resolved by launch_simulation)
    if args_cli.device is not None and not args_cli.distributed:
        agent_cfg["params"]["config"]["device"] = args_cli.device
        agent_cfg["params"]["config"]["device_name"] = args_cli.device

    # randomly sample a seed if seed = -1
    if args_cli.seed == -1:
        args_cli.seed = random.randint(0, 10000)
    agent_cfg["params"]["seed"] = args_cli.seed if args_cli.seed is not None else agent_cfg["params"]["seed"]

    # process distributed
    # env_cfg.sim.device is already resolved by launch_simulation().
    world_rank = 0
    if args_cli.distributed:
        agent_cfg["params"]["config"]["device"] = env_cfg.sim.device
        world_rank = int(os.getenv("RANK", "0"))

    # specify directory for logging experiments
    log_root_path = os.path.join("logs", "rl_games", agent_cfg["params"]["config"]["name"])
    log_root_path = os.path.abspath(log_root_path)
    print(f"[INFO] Logging experiment in directory: {log_root_path}")
    # specify directory for logging runs
    log_dir = agent_cfg["params"]["config"].get("full_experiment_name", datetime.now().strftime("%Y-%m-%d_%H-%M-%S"))
    # set directory into agent config
    # logging directory path: <train_dir>/<full_experiment_name>
    agent_cfg["params"]["config"]["train_dir"] = log_root_path
    agent_cfg["params"]["config"]["full_experiment_name"] = log_dir

    # multi-gpu training config
    if args_cli.distributed:
        agent_cfg["params"]["seed"] += int(os.getenv("RANK", "0"))
        agent_cfg["params"]["config"]["device"] = env_cfg.sim.device
        agent_cfg["params"]["config"]["device_name"] = env_cfg.sim.device
        agent_cfg["params"]["config"]["multi_gpu"] = True

    # max iterations
    if args_cli.max_iterations:
        agent_cfg["params"]["config"]["max_epochs"] = args_cli.max_iterations

    # dump the configuration into log-directory
    dump_yaml(os.path.join(log_root_path, log_dir, "params", "env.yaml"), env_cfg)
    dump_yaml(os.path.join(log_root_path, log_dir, "params", "agent.yaml"), agent_cfg)

    # read configurations about the agent-training
    rl_device = agent_cfg["params"]["config"]["device"]
    clip_obs = agent_cfg["params"]["env"].get("clip_observations", math.inf)
    clip_actions = agent_cfg["params"]["env"].get("clip_actions", math.inf)

    task_startup_time_begin = time.perf_counter_ns()

    # create isaac environment
    env = gym.make(args_cli.task, cfg=env_cfg, render_mode="rgb_array" if args_cli.video else None)
    # wrap for video recording
    if args_cli.video:
        video_kwargs = {
            "video_folder": os.path.join(log_root_path, log_dir, "videos"),
            "step_trigger": lambda step: step % args_cli.video_interval == 0,
            "video_length": args_cli.video_length,
            "disable_logger": True,
        }
        print("[INFO] Recording videos during training.")
        print_dict(video_kwargs, nesting=4)
        env = gym.wrappers.RecordVideo(env, **video_kwargs)

    # wrap around environment for rl-games
    env = RlGamesVecEnvWrapper(env, rl_device, clip_obs, clip_actions)

    task_startup_time_end = time.perf_counter_ns()

    # register the environment to rl-games registry
    # note: in agents configuration: environment name must be "rlgpu"
    vecenv.register(
        "IsaacRlgWrapper", lambda config_name, num_actors, **kwargs: RlGamesGpuEnv(config_name, num_actors, **kwargs)
    )
    env_configurations.register("rlgpu", {"vecenv_type": "IsaacRlgWrapper", "env_creator": lambda **kwargs: env})

    # set number of actors into agent config
    agent_cfg["params"]["config"]["num_actors"] = env.unwrapped.num_envs
    # always track the success metric; early-stop only if --check_success
    observer = RlGamesEarlyStopObserver(IsaacAlgoObserver(), **build_success_kwargs(args_cli))
    runner = Runner(observer)
    runner.load(agent_cfg)

    # set seed of the env
    env.seed(agent_cfg["params"]["seed"])
    # reset the agent and env
    runner.reset()

    # train the agent with continuous benchmark monitoring
    with BenchmarkMonitor(benchmark, interval=1.0):
        runner.run({"train": True, "play": False, "sigma": None})

    if world_rank == 0:
        # Final update after training completes
        benchmark.update_manual_recorders()

        # parse tensorboard file stats
        tensorboard_log_dir = os.path.join(log_root_path, log_dir, "summaries")
        log_data = parse_tf_logs(tensorboard_log_dir)

        # prepare RL timing dict
        rl_training_times = {
            "Environment only step time": log_data["performance/step_time"],
            "Environment + Inference step time": log_data["performance/step_inference_time"],
            "Environment + Inference + Policy update time": log_data["performance/rl_update_time"],
            "Environment only FPS": log_data["performance/step_fps"],
            "Environment + Inference FPS": log_data["performance/step_inference_fps"],
            "Environment + Inference + Policy update FPS": log_data["performance/step_inference_rl_update_fps"],
        }

        # log additional metrics to benchmark services
        log_app_start_time(benchmark, (app_start_time_end - app_start_time_begin) / 1e6)
        log_python_imports_time(benchmark, (imports_time_end - imports_time_begin) / 1e6)
        log_task_start_time(benchmark, (task_startup_time_end - task_startup_time_begin) / 1e6)
        log_scene_creation_time(benchmark, Timer.get_timer_info("scene_creation") * 1000)
        log_simulation_start_time(benchmark, Timer.get_timer_info("simulation_start") * 1000)
        log_total_start_time(benchmark, (task_startup_time_end - app_start_time_begin) / 1e6)
        log_runtime_step_times(benchmark, rl_training_times, compute_stats=True)
        log_rl_training_metrics(
            benchmark,
            log_data,
            reward_tag="rewards/iter",
            episode_length_tag="episode_lengths/iter",
            task=args_cli.task,
            workflow="rl_games",
            should_check_convergence=args_cli.check_convergence,
            reward_threshold=args_cli.reward_threshold,
            convergence_config=args_cli.convergence_config,
        )

        tracker = get_success_tracker(args_cli, observer.tracker, log_data)
        log_success(benchmark, tracker, framework_iteration_count=observer.framework_iteration_count)

        benchmark._finalize_impl()

    # close the simulator
    env.close()


if __name__ == "__main__":
    env_cfg, agent_cfg = resolve_task_config(args_cli.task, "rl_games_cfg_entry_point")

    app_start_time_begin = time.perf_counter_ns()
    with launch_simulation(env_cfg, args_cli):
        app_start_time_end = time.perf_counter_ns()
        main(env_cfg, agent_cfg, app_start_time_begin, app_start_time_end)