244 lines
10 KiB
Python
244 lines
10 KiB
Python
# Copyright (c) 2024-2025 Ziqi Fan
|
|
# SPDX-License-Identifier: Apache-2.0
|
|
|
|
# Copyright (c) 2024-2025, The Isaac Lab Project Developers.
|
|
# All rights reserved.
|
|
#
|
|
# SPDX-License-Identifier: Apache-2.0
|
|
|
|
"""Script to play a checkpoint if an RL agent from RSL-RL."""
|
|
|
|
"""Launch Isaac Sim Simulator first."""
|
|
|
|
import argparse
|
|
import os
|
|
import sys
|
|
|
|
from isaaclab.app import AppLauncher
|
|
|
|
# local imports
|
|
import cli_args # isort: skip
|
|
from utils import export_cts_policy_as_jit, export_cts_policy_as_onnx
|
|
|
|
# add argparse arguments
|
|
parser = argparse.ArgumentParser(description="Train an RL agent with RSL-RL.")
|
|
parser.add_argument("--video", action="store_true", default=False, help="Record videos during training.")
|
|
parser.add_argument("--video_length", type=int, default=int(1e9), help="Length of the recorded video (in steps).")
|
|
parser.add_argument(
|
|
"--disable_fabric", action="store_true", default=False, help="Disable fabric and use USD I/O operations."
|
|
)
|
|
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(
|
|
"--agent", type=str, default="rsl_rl_cfg_entry_point", help="Name of the RL agent configuration entry point."
|
|
)
|
|
parser.add_argument("--seed", type=int, default=None, help="Seed used for the environment")
|
|
parser.add_argument(
|
|
"--use_pretrained_checkpoint",
|
|
action="store_true",
|
|
help="Use the pre-trained checkpoint from Nucleus.",
|
|
)
|
|
parser.add_argument("--real-time", action="store_true", default=False, help="Run in real-time, if possible.")
|
|
parser.add_argument("--keyboard", action="store_true", default=False, help="Whether to use keyboard.")
|
|
parser.add_argument("--fix_commands", action="store_true", default=False, help="Fix the velocity commands.")
|
|
# append RSL-RL cli arguments
|
|
cli_args.add_rsl_rl_args(parser)
|
|
# append AppLauncher cli args
|
|
AppLauncher.add_app_launcher_args(parser)
|
|
# parse the arguments
|
|
args_cli, hydra_args = parser.parse_known_args()
|
|
# always enable cameras to record video
|
|
if args_cli.video:
|
|
args_cli.enable_cameras = True
|
|
|
|
# clear out sys.argv for Hydra
|
|
sys.argv = [sys.argv[0]] + hydra_args
|
|
|
|
# launch omniverse app
|
|
app_launcher = AppLauncher(args_cli)
|
|
simulation_app = app_launcher.app
|
|
|
|
"""Rest everything follows."""
|
|
|
|
import gymnasium as gym
|
|
import time
|
|
import torch
|
|
# from scripts.reinforcement_learning.utils import camera_follow
|
|
from rsl_rl.runners import DistillationRunner, OnPolicyRunner, OnPolicyRunnerCTS
|
|
|
|
from isaaclab.devices import Se2Keyboard, Se2KeyboardCfg
|
|
from isaaclab.envs import (
|
|
DirectMARLEnv,
|
|
DirectMARLEnvCfg,
|
|
DirectRLEnvCfg,
|
|
ManagerBasedRLEnvCfg,
|
|
multi_agent_to_single_agent,
|
|
)
|
|
from isaaclab.utils.assets import retrieve_file_path
|
|
from isaaclab.utils.dict import print_dict
|
|
# from isaaclab.utils.pretrained_checkpoint import get_published_pretrained_checkpoint
|
|
from isaaclab_rl.rsl_rl import RslRlBaseRunnerCfg, RslRlVecEnvWrapper, export_policy_as_jit, export_policy_as_onnx
|
|
from isaaclab_tasks.utils import get_checkpoint_path
|
|
from isaaclab_tasks.utils.hydra import hydra_task_config
|
|
import robot_lab.tasks # noqa: F401
|
|
|
|
def fix_commands(env_cfg: ManagerBasedRLEnvCfg | DirectRLEnvCfg | DirectMARLEnvCfg):
|
|
from isaaclab.managers import ObservationTermCfg as ObsTerm
|
|
num_envs = env_cfg.scene.num_envs
|
|
def fix(env):
|
|
return torch.tensor([1.0, 0.0, 0.0], device=env_cfg.sim.device).repeat(num_envs, 1)
|
|
env_cfg.observations.policy.velocity_commands = ObsTerm(func=fix)
|
|
env_cfg.observations.critic.velocity_commands = ObsTerm(func=fix)
|
|
|
|
@hydra_task_config(args_cli.task, args_cli.agent)
|
|
def main(env_cfg: ManagerBasedRLEnvCfg | DirectRLEnvCfg | DirectMARLEnvCfg, agent_cfg: RslRlBaseRunnerCfg):
|
|
"""Play with RSL-RL agent."""
|
|
# grab task name for checkpoint path
|
|
task_name = args_cli.task.split(":")[-1]
|
|
|
|
# override configurations with non-hydra CLI arguments
|
|
agent_cfg: RslRlBaseRunnerCfg = cli_args.update_rsl_rl_cfg(agent_cfg, args_cli)
|
|
env_cfg.scene.num_envs = args_cli.num_envs if args_cli.num_envs is not None else 64
|
|
|
|
# set the environment seed
|
|
# note: certain randomizations occur in the environment initialization so we set the seed here
|
|
env_cfg.seed = agent_cfg.seed
|
|
env_cfg.sim.device = args_cli.device if args_cli.device is not None else env_cfg.sim.device
|
|
|
|
# disable randomization for play
|
|
env_cfg.observations.policy.enable_corruption = False
|
|
# remove random pushing
|
|
env_cfg.events.randomize_apply_external_force_torque = None
|
|
env_cfg.events.push_robot = None
|
|
env_cfg.curriculum.command_levels_lin_vel = None
|
|
env_cfg.curriculum.command_levels_ang_vel = None
|
|
|
|
# specify directory for logging experiments
|
|
log_root_path = os.path.join("logs", "rsl_rl", agent_cfg.experiment_name)
|
|
log_root_path = os.path.abspath(log_root_path)
|
|
print(f"[INFO] Loading experiment from directory: {log_root_path}")
|
|
if args_cli.use_pretrained_checkpoint:
|
|
# resume_path = get_published_pretrained_checkpoint("rsl_rl", task_name)
|
|
# if not resume_path:
|
|
# print("[INFO] Unfortunately a pre-trained checkpoint is currently unavailable for this task.")
|
|
# return
|
|
raise NotImplementedError("Pre-trained checkpoint retrieval is disabled temporarily.")
|
|
elif args_cli.checkpoint:
|
|
resume_path = retrieve_file_path(args_cli.checkpoint)
|
|
else:
|
|
resume_path = get_checkpoint_path(log_root_path, agent_cfg.load_run, agent_cfg.load_checkpoint)
|
|
|
|
log_dir = os.path.dirname(resume_path)
|
|
|
|
# set the log directory for the environment (works for all environment types)
|
|
env_cfg.log_dir = log_dir
|
|
|
|
# fix velocity commands if specified
|
|
if args_cli.fix_commands:
|
|
fix_commands(env_cfg)
|
|
|
|
# create isaac environment
|
|
env = gym.make(args_cli.task, cfg=env_cfg, render_mode="rgb_array" if args_cli.video else None)
|
|
|
|
# convert to single-agent instance if required by the RL algorithm
|
|
if isinstance(env.unwrapped, DirectMARLEnv):
|
|
env = multi_agent_to_single_agent(env)
|
|
|
|
# wrap for video recording
|
|
if args_cli.video:
|
|
import imageio
|
|
video_path = os.path.join(log_dir, "videos", "play", time.strftime("%Y-%m-%d_%H-%M-%S") + ".mp4")
|
|
writer = imageio.get_writer(video_path, fps=int(1/env.unwrapped.step_dt))
|
|
# video_kwargs = {
|
|
# "video_folder": os.path.join(log_dir, "videos", "play"),
|
|
# "step_trigger": lambda step: step == 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) # Store all frames into list is slow and memory-consuming, use imageio
|
|
|
|
# wrap around environment for rsl-rl
|
|
env = RslRlVecEnvWrapper(env, clip_actions=agent_cfg.clip_actions)
|
|
|
|
print(f"[INFO]: Loading model checkpoint from: {resume_path}")
|
|
# load previously trained model
|
|
if agent_cfg.class_name == "OnPolicyRunner":
|
|
runner = OnPolicyRunner(env, agent_cfg.to_dict(), log_dir=None, device=agent_cfg.device)
|
|
elif agent_cfg.class_name == "DistillationRunner":
|
|
runner = DistillationRunner(env, agent_cfg.to_dict(), log_dir=None, device=agent_cfg.device)
|
|
elif agent_cfg.class_name == "OnPolicyRunnerCTS":
|
|
runner = OnPolicyRunnerCTS(env, agent_cfg.to_dict(), log_dir=log_dir, device=agent_cfg.device)
|
|
else:
|
|
raise ValueError(f"Unsupported runner class: {agent_cfg.class_name}")
|
|
runner.load(resume_path)
|
|
|
|
# obtain the trained policy for inference
|
|
policy = runner.get_inference_policy(device=env.unwrapped.device)
|
|
|
|
# extract the neural network module
|
|
# we do this in a try-except to maintain backwards compatibility.
|
|
try:
|
|
# version 2.3 onwards
|
|
policy_nn = runner.alg.policy
|
|
except AttributeError:
|
|
# version 2.2 and below
|
|
policy_nn = runner.alg.actor_critic
|
|
|
|
# extract the normalizer
|
|
if hasattr(policy_nn, "actor_obs_normalizer"):
|
|
normalizer = policy_nn.actor_obs_normalizer
|
|
elif hasattr(policy_nn, "student_obs_normalizer"):
|
|
normalizer = policy_nn.student_obs_normalizer
|
|
else:
|
|
normalizer = None
|
|
|
|
# export policy to onnx/jit
|
|
export_model_dir = os.path.join(os.path.dirname(resume_path), "exported")
|
|
if agent_cfg.class_name == "OnPolicyRunnerCTS":
|
|
export_cts_policy_as_jit(policy_nn, actor_obs_normalizer=policy_nn.actor_obs_normalizer, single_obs_normalizer=policy_nn.single_obs_normalizer, path=export_model_dir, filename="policy.pt")
|
|
export_cts_policy_as_onnx(policy_nn, actor_obs_normalizer=policy_nn.actor_obs_normalizer, single_obs_normalizer=policy_nn.single_obs_normalizer, path=export_model_dir, filename="policy.onnx")
|
|
else:
|
|
export_policy_as_jit(policy_nn, normalizer=normalizer, path=export_model_dir, filename="policy.pt")
|
|
export_policy_as_onnx(policy_nn, normalizer=normalizer, path=export_model_dir, filename="policy.onnx")
|
|
|
|
dt = env.unwrapped.step_dt
|
|
|
|
# env.unwrapped.eye = (1.1, 3.3, 0.9)
|
|
# reset environment
|
|
obs = env.get_observations()
|
|
timestep = 0
|
|
# simulate environment
|
|
while simulation_app.is_running():
|
|
start_time = time.time()
|
|
# run everything in inference mode
|
|
with torch.inference_mode():
|
|
# agent stepping
|
|
actions = policy(obs)
|
|
# env stepping
|
|
obs, _, dones, _ = env.step(actions)
|
|
# reset recurrent states for episodes that have terminated
|
|
policy_nn.reset(dones)
|
|
if args_cli.video:
|
|
writer.append_data(env.env.render())
|
|
# camera_follow(env)
|
|
|
|
# time delay for real-time evaluation
|
|
sleep_time = dt - (time.time() - start_time)
|
|
if args_cli.real_time and sleep_time > 0:
|
|
time.sleep(sleep_time)
|
|
|
|
# close the simulator
|
|
env.close()
|
|
|
|
# close the video writer
|
|
if args_cli.video:
|
|
writer.close()
|
|
|
|
if __name__ == "__main__":
|
|
# run the main function
|
|
main()
|
|
# close sim app
|
|
simulation_app.close()
|