diff --git a/src/box/moblARMslndex.py b/src/box/moblARMslndex.py
new file mode 100644
index 0000000000..950f646641
--- /dev/null
+++ b/src/box/moblARMslndex.py
@@ -0,0 +1,270 @@
+"""
+MoblArmsIndex 生物力学模型 - 修改版
+这是基于 MoBL ARMS 模型的 MuJoCo 实现,专门用于手指指向任务
+"""
+
+import sys
+import os
+
+# ================ 动态导入 BaseBMModel ================
+try:
+ # 尝试1: 从原始相对位置导入
+ from ..base import BaseBMModel
+ print("✓ 使用相对导入: from ..base import BaseBMModel")
+except ImportError:
+ try:
+ # 尝试2: 从绝对路径导入(假设在 uitb 包中)
+ from uitb.base import BaseBMModel
+ print("✓ 使用绝对导入: from uitb.base import BaseBMModel")
+ except ImportError:
+ try:
+ # 尝试3: 从常见路径导入
+ import sys
+ import os
+
+ # 将项目根目录添加到路径
+ current_dir = os.path.dirname(os.path.abspath(__file__))
+ project_root = os.path.dirname(os.path.dirname(current_dir))
+ sys.path.insert(0, project_root)
+
+ from base import BaseBMModel
+ print(f"✓ 使用项目根目录导入: from base import BaseBMModel (路径: {project_root})")
+ except ImportError as e:
+ # 尝试4: 搜索并动态导入
+ print("⚠ 无法自动导入 BaseBMModel,尝试动态查找...")
+
+ # 搜索 base.py 文件
+ base_path = None
+ search_dirs = [
+ os.path.dirname(os.path.dirname(os.path.abspath(__file__))), # 上一级目录
+ os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))), # 上上级目录
+ os.path.join(os.path.dirname(__file__), '..'), # 相对路径
+ ]
+
+ for search_dir in search_dirs:
+ possible_path = os.path.join(search_dir, 'base.py')
+ if os.path.exists(possible_path):
+ base_path = possible_path
+ break
+
+ if base_path:
+ # 动态导入
+ import importlib.util
+ spec = importlib.util.spec_from_file_location("base_module", base_path)
+ base_module = importlib.util.module_from_spec(spec)
+ sys.modules[spec.name] = base_module
+ spec.loader.exec_module(base_module)
+ BaseBMModel = base_module.BaseBMModel
+ print(f"✓ 动态导入成功: {base_path}")
+ else:
+ # 如果都失败,创建简化的 BaseBMModel
+ print("⚠ 无法找到 BaseBMModel,使用简化版本")
+
+ class BaseBMModel:
+ """简化的 BaseBMModel 基类"""
+ def __init__(self, model, data, **kwargs):
+ self.model = model
+ self.data = data
+
+ def _update(self, model, data):
+ """更新方法 - 子类应重写"""
+ pass
+
+ @classmethod
+ def _get_floor(cls):
+ """获取地板 - 子类应重写"""
+ return None
+
+ BaseBMModel = BaseBMModel
+ print("✓ 使用简化版本的 BaseBMModel")
+
+# ================ 继续原来的代码 ================
+import numpy as np
+import mujoco
+
+
+class MoblArmsIndex(BaseBMModel):
+ """
+ 基于 MoBL ARMS 模型的生物力学模型
+
+ 来源:
+ - 原始 OpenSim 模型: https://simtk.org/frs/?group_id=657
+ - MuJoCo 转换工具: https://github.com/aikkala/O2MConverter
+
+ 说明:
+ 此模型与 uitb/bm_models/mobl_arms 中的模型相同,
+ 但食指处于弯曲状态并包含力传感器。
+ """
+
+ def __init__(self, model, data, **kwargs):
+ """
+ 初始化 MoblArmsIndex 模型
+
+ 参数:
+ model: MuJoCo 模型实例
+ data: MuJoCo 数据实例
+ **kwargs: 额外参数
+ - shoulder_variant: 肩部变体,可选 "none" (默认) 或 "patch-v1"
+ """
+ super().__init__(model, data, **kwargs)
+
+ # 设置肩部变体
+ # 使用 "none" 作为默认值
+ # 使用 "patch-v1" 可以获得更合理的运动外观(未经全面测试)
+ self.shoulder_variant = kwargs.get("shoulder_variant", "none")
+
+ print(f"✓ MoblArmsIndex 初始化完成,肩部变体: {self.shoulder_variant}")
+
+ def _update(self, model, data):
+ """
+ 更新模型状态
+
+ 此方法在每个时间步被调用,用于更新肩部约束。
+
+ 参数:
+ model: MuJoCo 模型实例
+ data: MuJoCo 数据实例
+ """
+ # 更新肩部等式约束
+ if self.shoulder_variant.startswith("patch"):
+ # 约束1: shoulder1_r2_con 约束的数据更新
+ model.equality("shoulder1_r2_con").data[1] = \
+ -((np.pi - 2 * data.joint('shoulder_elv').qpos) / np.pi)
+
+ # patch-v2 变体有额外的约束
+ if self.shoulder_variant == "patch-v2":
+ # 动态调整肩部旋转范围
+ data.joint('shoulder_rot').range[:] = \
+ np.array([-np.pi / 2, np.pi / 9]) - \
+ 2 * np.min((data.joint('shoulder_elv').qpos,
+ np.pi - data.joint('shoulder_elv').qpos)) / np.pi \
+ * data.joint('elv_angle').qpos
+
+ # 执行前向计算,更新物理状态
+ mujoco.mj_forward(model, data)
+
+ # 打印调试信息(可选)
+ if hasattr(self, 'debug') and self.debug:
+ print(f"更新肩部约束: shoulder_elv={data.joint('shoulder_elv').qpos:.3f}, "
+ f"约束值={model.equality('shoulder1_r2_con').data[1]:.3f}")
+
+ @classmethod
+ def _get_floor(cls):
+ """
+ 获取地板配置
+
+ 此模型不包含地板,返回 None。
+
+ 返回:
+ None: 表示此模型没有地板
+ """
+ return None
+
+ def get_force_sensor_data(self, data, sensor_name="index_force_sensor"):
+ """
+ 获取力传感器数据
+
+ 参数:
+ data: MuJoCo 数据实例
+ sensor_name: 传感器名称,默认为 "index_force_sensor"
+
+ 返回:
+ numpy.ndarray: 传感器数据,如果没有找到传感器则返回 None
+ """
+ try:
+ sensor_id = self.model.sensor(sensor_name).id
+ return data.sensordata[sensor_id]
+ except Exception as e:
+ if hasattr(self, 'debug') and self.debug:
+ print(f"⚠ 无法获取力传感器数据: {e}")
+ return None
+
+ def set_debug_mode(self, debug=True):
+ """
+ 设置调试模式
+
+ 参数:
+ debug: 是否启用调试模式
+ """
+ self.debug = debug
+
+ def __str__(self):
+ """
+ 返回模型的字符串表示
+
+ 返回:
+ str: 模型描述
+ """
+ return (f"MoblArmsIndex 模型 (肩部变体: {self.shoulder_variant})\n"
+ f"描述: 基于 MoBL ARMS 的上肢模型,食指弯曲并包含力传感器")
+
+
+# ================ 测试代码 ================
+if __name__ == "__main__":
+ """
+ 直接运行此文件的测试代码
+ """
+ print("=" * 60)
+ print("MoblArmsIndex 模型测试")
+ print("=" * 60)
+
+ # 测试导入
+ print("1. 测试导入和类定义:")
+ print(f" 类名: {MoblArmsIndex.__name__}")
+ print(f" 基类: {MoblArmsIndex.__bases__[0].__name__}")
+ print(f" 文档: {MoblArmsIndex.__doc__.strip().split('\n')[0]}")
+
+ # 测试创建实例(需要 MuJoCo 模型)
+ try:
+ import mujoco
+ import numpy as np
+
+ print("\n2. 测试创建实例:")
+
+ # 创建一个简单的 MuJoCo 模型用于测试
+ xml_string = """
+
+
+
+
+
+
+
+
+ """
+
+ model = mujoco.MjModel.from_xml_string(xml_string)
+ data = mujoco.MjData(model)
+
+ # 创建模型实例
+ bm_model = MoblArmsIndex(model, data, shoulder_variant="none")
+
+ print(f" ✓ 成功创建 MoblArmsIndex 实例")
+ print(f" 肩部变体: {bm_model.shoulder_variant}")
+
+ # 测试更新方法
+ print("\n3. 测试更新方法:")
+ bm_model._update(model, data)
+ print(" ✓ _update 方法执行成功")
+
+ # 测试获取地板
+ print("\n4. 测试获取地板:")
+ floor = bm_model._get_floor()
+ print(f" 地板配置: {floor}")
+
+ # 测试字符串表示
+ print("\n5. 测试字符串表示:")
+ print(f" {bm_model}")
+
+ print("\n" + "=" * 60)
+ print("所有测试通过!")
+ print("=" * 60)
+
+ except Exception as e:
+ print(f"\n✗ 测试过程中出现错误: {type(e).__name__}: {e}")
+ import traceback
+ traceback.print_exc()
+
+ print("\n" + "=" * 60)
+ print("注意: 缺少 MuJoCo 模型文件,但导入测试成功")
+ print("=" * 60)
diff --git a/src/box/perception/moblARMslndex.py b/src/box/perception/moblARMslndex.py
new file mode 100644
index 0000000000..299bbb4d67
--- /dev/null
+++ b/src/box/perception/moblARMslndex.py
@@ -0,0 +1,372 @@
+"""
+MoblArmsIndex 简化版本 - 只生成人体上肢模型运动视频
+完全修复版
+"""
+
+import numpy as np
+import mujoco
+import imageio
+import os
+import subprocess
+import sys
+
+def install_ffmpeg():
+ """安装必要的视频编码依赖"""
+ try:
+ # 尝试导入 imageio-ffmpeg
+ import imageio_ffmpeg
+ print("✓ imageio-ffmpeg 已安装")
+ return True
+ except ImportError:
+ print("正在安装 imageio-ffmpeg...")
+ try:
+ subprocess.check_call([sys.executable, "-m", "pip", "install", "imageio[ffmpeg]"])
+ print("✓ imageio[ffmpeg] 安装成功")
+ return True
+ except Exception as e:
+ print(f"⚠ 安装失败: {e}")
+ print("将使用 GIF 格式替代 MP4")
+ return False
+
+# ================ 简化的 BaseBMModel 类 ================
+class BaseBMModel:
+ """简化的 BaseBMModel 基类"""
+ def __init__(self, model, data, **kwargs):
+ self.model = model
+ self.data = data
+
+ def _update(self, model, data):
+ """更新方法 - 子类应重写"""
+ pass
+
+# ================ MoblArmsIndex 模型 ================
+class MoblArmsIndex(BaseBMModel):
+ """
+ 基于 MoBL ARMS 模型的生物力学模型
+ 只专注于生成上肢运动视频
+ """
+
+ def __init__(self, model, data, **kwargs):
+ super().__init__(model, data, **kwargs)
+ print("✓ MoblArmsIndex 上肢模型初始化完成")
+
+ def _update(self, model, data):
+ """更新模型状态 - 简化版本"""
+ # 这里可以添加任何必要的模型更新逻辑
+ pass
+
+# ================ 主程序:只生成视频 ================
+def main():
+ print("=" * 60)
+ print("人体上肢模型运动视频生成器")
+ print("=" * 60)
+
+ # 检查并安装必要依赖
+ has_ffmpeg = install_ffmpeg()
+
+ # 1. 创建更真实的上肢XML模型(修复了所有问题)
+ xml_string = """
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+ """
+
+ # 2. 加载模型
+ try:
+ model = mujoco.MjModel.from_xml_string(xml_string)
+ data = mujoco.MjData(model)
+ print("✓ MuJoCo 模型加载成功")
+ print(f" 关节数: {model.njnt}")
+ print(f" 几何体数: {model.ngeom}")
+ print(f" 执行器数: {model.nu}")
+ except Exception as e:
+ print(f"❌ 模型加载失败: {e}")
+ return
+
+ # 3. 创建生物力学模型实例
+ bm_model = MoblArmsIndex(model, data)
+
+ # 4. 设置仿真参数
+ fps = 30
+ duration = 5 # 5秒仿真
+ total_steps = int(duration / model.opt.timestep)
+
+ print(f"\n仿真参数:")
+ print(f" 时长: {duration}秒")
+ print(f" 时间步长: {model.opt.timestep}")
+ print(f" 总步数: {total_steps}")
+ print(f" 视频帧率: {fps} FPS")
+
+ # 5. 设置渲染器(使用较小的分辨率避免问题)
+ try:
+ # 创建渲染器(使用640x480避免离屏缓冲区问题)
+ renderer = mujoco.Renderer(model, height=480, width=640)
+
+ # 设置相机视角
+ camera = mujoco.MjvCamera()
+ camera.azimuth = 30 # 水平角度
+ camera.elevation = -20 # 垂直角度
+ camera.distance = 2.5 # 距离
+ camera.lookat = np.array([0.2, 0, 1.2]) # 看向点
+
+ # 初始渲染
+ renderer.update_scene(data, camera=camera)
+ print("✓ 渲染器创建成功")
+ except Exception as e:
+ print(f"❌ 渲染器创建失败: {e}")
+ return
+
+ # 6. 运行仿真并捕获视频帧
+ frames = []
+
+ print("\n开始仿真并录制视频...")
+
+ for step in range(total_steps):
+ # 计算时间(用于控制运动)
+ t = step * model.opt.timestep
+
+ # 协同运动控制
+ # 肩部运动(缓慢的波浪)
+ shoulder_pitch = 0.3 * np.sin(2 * np.pi * 0.3 * t) + 0.1
+ shoulder_roll = 0.2 * np.sin(2 * np.pi * 0.2 * t + 0.5)
+
+ # 肘部运动
+ elbow = 0.6 * np.sin(2 * np.pi * 0.4 * t) - 1.0
+
+ # 手腕运动
+ wrist = 0.1 * np.sin(2 * np.pi * 0.6 * t)
+
+ # 手指运动序列
+ if t < 1:
+ # 初始:手指弯曲
+ index_angle = -0.4
+ thumb_angle = 0.1
+ elif t < 3:
+ # 指向阶段
+ phase = (t - 1) / 2
+ index_angle = 0.1 * np.sin(2 * np.pi * 1.0 * t) - 0.1
+ thumb_angle = 0.15 * np.sin(2 * np.pi * 0.8 * t)
+ else:
+ # 返回阶段
+ phase = (t - 3) / 2
+ index_angle = -0.4 * (1 - phase)
+ thumb_angle = 0.1 * (1 - phase)
+
+ # 应用控制信号
+ data.ctrl[0] = shoulder_pitch # 肩部俯仰
+ data.ctrl[1] = shoulder_roll # 肩部滚动
+ data.ctrl[2] = elbow # 肘部
+ data.ctrl[3] = wrist # 手腕
+ data.ctrl[4] = index_angle # 食指
+ data.ctrl[5] = thumb_angle # 拇指
+
+ # 步进仿真
+ mujoco.mj_step(model, data)
+
+ # 定期捕获帧
+ if step % 2 == 0: # 每2步捕获一帧,获得15fps(GIF比较流畅)
+ try:
+ # 动态调整相机视角
+ if t < 2:
+ camera.azimuth = 30 + 15 * t
+ elif t < 4:
+ camera.elevation = -20 + 10 * np.sin(t * 0.5)
+
+ # 更新场景并渲染
+ renderer.update_scene(data, camera=camera)
+ frame = renderer.render()
+ frames.append(frame)
+ except:
+ # 如果失败,使用默认视角
+ try:
+ renderer.update_scene(data)
+ frame = renderer.render()
+ frames.append(frame)
+ except:
+ pass
+
+ # 显示进度
+ if step % 100 == 0:
+ progress = (step / total_steps) * 100
+ print(f"进度: {progress:.1f}%", end='\r')
+
+ print(f"进度: 100.0% - 仿真完成!")
+
+ # 7. 生成视频文件
+ if frames:
+ try:
+ # 确定保存路径(桌面)
+ desktop = os.path.join(os.path.expanduser("~"), "Desktop")
+ if not os.path.exists(desktop):
+ desktop = os.path.dirname(os.path.abspath(__file__))
+
+ # 首先尝试生成MP4(如果有FFMPEG)
+ if has_ffmpeg:
+ try:
+ video_path = os.path.join(desktop, "human_arm_motion.mp4")
+
+ # 保存MP4视频
+ print(f"\n生成MP4视频中...")
+ imageio.mimsave(
+ video_path,
+ frames,
+ fps=15, # GIF帧率较低
+ codec='libx264',
+ quality=8
+ )
+ print(f"✓ MP4视频已保存: {video_path}")
+ print(f" 文件大小: {os.path.getsize(video_path)/1024/1024:.1f} MB")
+
+ except Exception as mp4_error:
+ print(f"⚠ MP4生成失败: {mp4_error}")
+ # 回退到GIF
+ has_ffmpeg = False
+
+ # 如果没有FFMPEG或MP4生成失败,生成GIF
+ if not has_ffmpeg:
+ gif_path = os.path.join(desktop, "human_arm_motion.gif")
+
+ print(f"\n生成GIF动画中...")
+ # 保存为GIF(质量稍低但兼容性好)
+ imageio.mimsave(
+ gif_path,
+ frames,
+ fps=15, # GIF帧率较低
+ subrectangles=True # 优化GIF大小
+ )
+ print(f"✓ GIF动画已保存: {gif_path}")
+ print(f" 文件大小: {os.path.getsize(gif_path)/1024:.1f} KB")
+
+ except Exception as e:
+ print(f"❌ 文件保存失败: {e}")
+ else:
+ print("⚠ 没有捕获到视频帧")
+
+ print("\n" + "=" * 60)
+ print("✅ 任务完成!")
+ print("=" * 60)
+
+ # 8. 打开文件所在目录(仅Windows)
+ if 'video_path' in locals() and os.path.exists(video_path):
+ print(f"📁 MP4视频文件: {video_path}")
+ try:
+ os.startfile(os.path.dirname(video_path))
+ print("📂 已打开文件所在目录")
+ except:
+ pass
+ elif 'gif_path' in locals() and os.path.exists(gif_path):
+ print(f"📁 GIF动画文件: {gif_path}")
+ try:
+ os.startfile(os.path.dirname(gif_path))
+ print("📂 已打开文件所在目录")
+ except:
+ pass
+
+# ================ 运行主程序 ================
+if __name__ == "__main__":
+ # 检查必要依赖
+ try:
+ import mujoco
+ import numpy as np
+ import imageio
+ print("✓ 核心库已安装")
+ except ImportError as e:
+ print(f"❌ 缺少必要的库: {e}")
+ print("\n请安装以下依赖:")
+ print("pip install mujoco numpy imageio")
+ exit(1)
+
+ try:
+ main()
+ except KeyboardInterrupt:
+ print("\n⚠ 用户中断")
+ except Exception as e:
+ print(f"\n❌ 运行出错: {type(e).__name__}: {e}")
+ import traceback
+ traceback.print_exc()
+
\ No newline at end of file
diff --git a/src/box/trainer.py b/src/box/trainer.py
new file mode 100644
index 0000000000..7170f873d6
--- /dev/null
+++ b/src/box/trainer.py
@@ -0,0 +1,121 @@
+import sys
+import os, shutil
+import logging
+from datetime import datetime
+
+import wandb
+from wandb.integration.sb3 import WandbCallback
+
+from uitb.simulator import Simulator
+from uitb.utils.functions import output_path, timeout_input
+
+from stable_baselines3.common.save_util import load_from_zip_file
+
+import argparse
+
+if __name__=="__main__":
+
+ parser = argparse.ArgumentParser(description='Train an agent.')
+ parser.add_argument('config_file_path', type=str,
+ help='path to the config file')
+ parser.add_argument('--checkpoint', type=str, default=None,
+ help='filename of a specific checkpoint to resume training at '
+ '(default: None, start training from the scratch)')
+ parser.add_argument('--resume', action='store_true', help='resume at latest checkpoint')
+ parser.add_argument('--eval', type=int, default=None, const=400000, nargs='?', help='run and store evaluations at a specific frequency (every ``eval`` timestep)')
+ parser.add_argument('--eval_info_keywords', type=str, nargs='*', default=[], help='additional keys of ``info`` dict that should be logged at the end of each evaluation episode')
+ args = parser.parse_args()
+
+ # Get config file path
+ config_file_path = args.config_file_path
+
+ # Build the simulator
+ simulator_folder = Simulator.build(config_file_path)
+
+ # Initialise
+ simulator = Simulator.get(simulator_folder)
+
+ # Get the config
+ config = simulator.config
+
+ # Get simulator name
+ name = config.get("simulator_name", None)
+
+ # Get checkpoint dir
+ checkpoint_dir = os.path.join(simulator._simulator_folder, 'checkpoints')
+
+ # Restore wandb run_id from checkpoint if available
+ checkpoint = args.checkpoint
+ resume_training = args.resume or (checkpoint is not None)
+ if resume_training:
+ if os.path.isdir(checkpoint_dir):
+ existing_checkpoints = [os.path.join(checkpoint_dir, f) for f in os.listdir(checkpoint_dir) if
+ os.path.isfile(os.path.join(checkpoint_dir, f))]
+ else:
+ raise FileNotFoundError(f"Checkpoint directory does not exist: {checkpoint_dir}\nTry to run without --checkpoint or --resume.")
+
+ if checkpoint is not None:
+ checkpoint_path = os.path.join(checkpoint_dir, checkpoint)
+ if not os.path.isfile(checkpoint_path):
+ raise FileNotFoundError(f"Checkpoint does not exist: {checkpoint_path}")
+ else:
+ assert len(existing_checkpoints) > 0, f"There are no checkpoints found in checkpoint directory: {checkpoint_dir}\n" \
+ f"Maybe existing checkpoints were moved to a backup directory, " \
+ f"which can be renamed to 'checkpoints/' to resume training."
+ # # find largest checkpoint
+ # checkpoint_path = sorted(existing_checkpoints, key=lambda f: int(f.split("_steps")[0].split("_")[-1]))[-1]
+ # find latest checkpoint
+ checkpoint_path = sorted(existing_checkpoints, key=os.path.getctime)[-1]
+ try:
+ _data, _, _ = load_from_zip_file(checkpoint_path)
+ wandb_id = _data["policy_kwargs"]["wandb_id"]
+ print(f"Resume wandb run {wandb_id} starting at checkpoint {checkpoint_path}.")
+ except Exception:
+ logging.warning("Cannot reliably identify wandb run id. Will resume training, but with new wandb instance and with step counter reset to zero.")
+ wandb_id = None
+ else:
+ checkpoint_path = None
+ wandb_id = None
+ # Backup existing checkpoint directory
+ if os.path.isdir(checkpoint_dir):
+ existing_checkpoints = [os.path.join(checkpoint_dir, f) for f in os.listdir(checkpoint_dir) if
+ os.path.isfile(os.path.join(checkpoint_dir, f))]
+ if len(existing_checkpoints) > 0:
+ last_checkpoint_time = max([os.path.getctime(f) for f in existing_checkpoints])
+ last_checkpoint_time = datetime.fromtimestamp(last_checkpoint_time).strftime('%Y%m%d_%H%M%S')
+ checkpoint_dir_backup = os.path.join(simulator._simulator_folder, f'checkpoints_{last_checkpoint_time}')
+ shutil.move(checkpoint_dir, checkpoint_dir_backup)
+
+ if name is None:
+ # Ask user to name this run
+ name = timeout_input("Give a name for this run. Input empty string or wait for 30 seconds for a random name.",
+ timeout=30, default="")
+ config["simulator_name"] = name.replace("-", "_")
+
+ # Get project name
+ project = config.get("project", "uitb")
+
+ # Prepare evaluation by storing custom log tags
+ with_evaluation = args.eval is not None
+ eval_freq = args.eval
+ eval_info_keywords = tuple(args.eval_info_keywords)
+
+ # Initialise wandb
+ if wandb_id is None:
+ wandb_id = wandb.util.generate_id()
+ run = wandb.init(id=wandb_id, resume="allow", project=project, name=name, config=config, sync_tensorboard=True, save_code=True, dir=output_path())
+
+ # Initialise RL model
+ rl_cls = simulator.get_class("rl", config["rl"]["algorithm"])
+ rl_model = rl_cls(simulator, checkpoint_path=checkpoint_path, wandb_id=wandb_id)
+
+ # Haven't figured out how to periodically save rl in wandb; this is currently done inside the rl_model class
+ # TODO this doesn't seem to work; do the files need to be in wandb.run.dir?
+ #wandb.save(os.path.join(model_folder, run.name, 'checkpoints', "model_*_steps.zip"),
+ # base_path=os.path.join(model_folder, run.name, 'checkpoints'))
+
+ # Start the training
+ # rl_model.learn(WandbCallback(verbose=2))
+ rl_model.learn(WandbCallback(verbose=2),
+ with_evaluation=with_evaluation, eval_freq=eval_freq, eval_info_keywords=eval_info_keywords)
+ run.finish()