feat: initial mario rl mvp

2026-02-12 18:54:06 +08:00
commit d23de69b9a
9 changed files with 1470 additions and 0 deletions
--- a/mario-rl-mvp/README.md
+++ b/mario-rl-mvp/README.md
@@ -0,0 +1,339 @@
+# Mario RL MVP (macOS Apple Silicon)
+
+最小可运行工程：使用像素输入 + 传统 CNN policy（`stable-baselines3` PPO）训练 `gym-super-mario-bros / nes-py` 智能体，不使用大语言模型。
+
+## 1. 项目结构
+
+```text
+mario-rl-mvp/
+  src/
+    env.py
+    train_ppo.py
+    eval.py
+    record_video.py
+    utils.py
+  artifacts/
+    models/
+    videos/
+    logs/
+  requirements.txt
+  README.md
+```
+
+## 2. 环境准备（macOS / Apple Silicon）
+
+建议 Python 3.9+（本机默认 `python3` 即可）。
+
+```bash
+cd /Users/roog/Work/FNT/SpMr/mario-rl-mvp
+python3 -m venv .venv
+source .venv/bin/activate
+python -m pip install --upgrade pip setuptools wheel
+pip install -r requirements.txt
+```
+
+可选系统依赖（用于 ffmpeg 转码与潜在 SDL 兼容）：
+
+```bash
+brew install ffmpeg sdl2
+```
+
+## 3. 一条命令开始训练
+
+默认 CPU 训练（如果检测到可用且稳定的 MPS，会自动尝试启用，否则自动回退 CPU）：
+
+```bash
+python -m src.train_ppo
+```
+
+常用覆盖参数：
+
+```bash
+python -m src.train_ppo \
+  --total-timesteps 1000000 \
+  --n-envs 4 \
+  --save-freq 50000 \
+  --env-id SuperMarioBros-1-1-v0 \
+  --movement right_only \
+  --seed 42
+```
+
+从已有 checkpoint 初始化后继续训练（可同时改超参数，如 `--ent-coef`）：
+
+```bash
+python -m src.train_ppo \
+  --init-model-path artifacts/models/latest_model.zip \
+  --total-timesteps 500000 \
+  --ent-coef 0.02 \
+  --learning-rate 1e-4
+```
+
+如果切换了动作空间（例如 checkpoint 是 `right_only`，当前想用 `simple`），可用部分加载：
+
+```bash
+python -m src.train_ppo \
+  --init-model-path artifacts/models/latest_model.zip \
+  --allow-partial-init \
+  --movement simple \
+  --total-timesteps 300000
+```
+
+### 3.1 从已有模型继续训练（`--init-model-path`）
+
+- 用途：加载已有 `.zip` 权重后继续训练，适合“不中断实验目标但调整探索参数”。
+- 常见场景：当前策略陷入局部最优（例如 `approx_kl` 和 `policy_gradient_loss` 长期接近 0），希望从已有模型继续探索。
+- 注意：这不是“热更新”，仍然需要停止当前训练进程后用新命令重启。
+
+```bash
+python -m src.train_ppo \
+  --init-model-path artifacts/models/latest_model.zip \
+  --ent-coef 0.02 \
+  --learning-rate 1e-4 \
+  --total-timesteps 300000
+```
+
+训练输出：
+- stdout：PPO 训练日志
+- TensorBoard：`artifacts/logs/<run_name>/tb/`
+- checkpoint：`artifacts/models/<run_name>/ppo_mario_ckpt_*.zip`
+- final model：`artifacts/models/<run_name>/final_model.zip`
+- latest 指针：`artifacts/models/latest_model.zip` + `LATEST_MODEL.txt`
+
+启动 TensorBoard：
+
+```bash
+tensorboard --logdir artifacts/logs --port 6006
+```
+
+## 3.2 训练日志字段速查（PPO）
+
+训练时你会看到类似：
+
+```text
+| rollout/ep_len_mean | rollout/ep_rew_mean | ... |
+| train/approx_kl     | train/entropy_loss  | ... |
+```
+
+下面是常用字段的含义（对应你贴出来那组）：
+
+- `rollout/ep_len_mean`：最近一批 episode 的平均步数。越大不一定越好，要结合 reward 一起看。
+- `rollout/ep_rew_mean`：最近一批 episode 的平均回报。通常越高越好。
+- `time/fps`：训练吞吐（每秒环境步数），只代表速度，不代表策略质量。
+- `time/iterations`：第几次 rollout + update 循环。
+- `time/time_elapsed`：训练已运行的秒数。
+- `time/total_timesteps`：累计环境交互步数（达到你设定的 `--total-timesteps` 会停止）。
+- `train/approx_kl`：新旧策略差异大小。太大说明更新过猛；接近 0 说明几乎没在更新。极小负数通常是数值误差，可当作 0。
+- `train/clip_fraction`：有多少样本触发 PPO clipping。长期为 0 且 KL 也接近 0，常见于“策略基本不再更新”。
+- `train/clip_range`：PPO 的 clipping 阈值（默认 0.2）。
+- `train/entropy_loss`：探索强度指标。绝对值越接近 0，策略越确定、探索越少。
+- `train/explained_variance`：价值网络对回报的解释度，越接近 1 越好，接近 0 说明 value 还不稳。
+- `train/learning_rate`：优化器步长（参数更新幅度），不是硬件速度。
+- `train/loss`：总损失（由多个部分组成），主要看趋势，不单看绝对值。
+- `train/policy_gradient_loss`：策略网络的更新信号。长期接近 0 可能表示 actor 更新很弱。
+- `train/value_loss`：价值网络误差。过大通常代表 critic 拟合还不稳定。
+
+### 快速判断（实用版）
+
+- `ep_rew_mean` / `avg_max_x_pos` 持续上升：一般在变好。
+- `approx_kl≈0` + `clip_fraction=0` + `policy_gradient_loss≈0`：大概率卡住（更新几乎停了）。
+- `entropy_loss` 绝对值太小且长期不变：探索不足，可尝试提高 `--ent-coef`。
+
+## 4. 评估模型
+
+加载最新模型，跑 N 个 episode，输出平均指标：
+
+```bash
+python -m src.eval --episodes 5
+```
+
+可指定模型：
+
+```bash
+python -m src.eval --model-path artifacts/models/latest_model.zip --episodes 10
+```
+
+注意：`eval.py` 默认 `--movement auto`，会按模型动作维度自动匹配 `right_only/simple`，避免动作空间不一致导致 `KeyError`。
+
+输出指标包括：
+- `avg_reward`
+- `avg_max_x_pos`
+- `clear_rate`（`flag_get=True` 的比例）
+
+
+查看步数
+
+```php
+unzip -p artifacts/models/latest_model.zip data | rg '"num_timesteps"|"_total_timesteps"|"_tensorboard_log"'
+```
+
+num_timesteps = 151552
+_total_timesteps = 150000
+
+## 5. 录制回放视频（无窗口/headless）
+
+默认录制约 10 秒 mp4 到 `artifacts/videos/`：
+
+```bash
+python -m src.record_video --duration-sec 10 --fps 30
+```
+
+可指定输出路径：
+
+```bash
+python -m src.record_video --output artifacts/videos/demo.mp4 --duration-sec 10
+```
+
+注意：`record_video.py` 默认 `--movement auto`，会按模型自动匹配动作空间。
+
+实现方式：
+- 使用 `render_mode=rgb_array`，无需打开窗口
+- 默认通过 `imageio + ffmpeg` 输出 mp4
+- 若 mp4 写入失败，会自动降级保存帧序列（PNG），并打印 ffmpeg 转码命令
+
+## 6. 动作空间选择说明
+
+默认 `RIGHT_ONLY`，原因：
+- 动作更少，探索空间更小，MVP 更快收敛到“向右推进”策略
+- 适合先验证训练闭环
+
+可切到 `SIMPLE_MOVEMENT`（动作更丰富）：
+
+```bash
+python -m src.train_ppo --movement simple
+```
+
+## 7. 预处理与奖励
+
+默认预处理链路：
+- 跳帧：`frame_skip=4`
+- 灰度：`RGB -> Gray`
+- 缩放：`84x84`
+- 帧堆叠：`4`
+- 通道布局：`CHW`（兼容 `CnnPolicy`）
+
+奖励：
+- `--reward-mode raw`：原始奖励（默认）
+- `--reward-mode clip`：裁剪奖励 `sign(reward)`（等价于旧参数 `--clip-reward`）
+- `--reward-mode progress`：奖励塑形模式，额外包含：
+  - 前进增益奖励（`--progress-scale`）
+  - 死亡惩罚（`--death-penalty`）
+  - 通关奖励（`--flag-bonus`）
+  - 原地卡住惩罚（`--stall-penalty` + `--stall-steps`）
+  - 后退惩罚（`--backward-penalty-scale`）
+
+使用裁剪奖励：
+
+```bash
+python -m src.train_ppo --reward-mode clip
+```
+
+针对“卡在固定位置（如 x=314 撞蘑菇）”的推荐续训命令：
+
+```bash
+python -m src.train_ppo \
+  --init-model-path artifacts/models/latest_model.zip \
+  --allow-partial-init \
+  --reward-mode progress \
+  --movement simple \
+  --ent-coef 0.04 \
+  --learning-rate 1e-4 \
+  --n-steps 512 \
+  --gamma 0.995 \
+  --death-penalty -120 \
+  --stall-penalty 0.2 \
+  --stall-steps 20 \
+  --backward-penalty-scale 0.03 \
+  --milestone-bonus 2.0 \
+  --no-progress-terminate-steps 80 \
+  --no-progress-terminate-penalty 30 \
+  --total-timesteps 150000
+```
+
+## 8. 常见问题排查
+
+### 8.1 `pip install` 失败（nes-py/gym-super-mario-bros 编译问题）
+
+先安装工具链并重试：
+
+```bash
+xcode-select --install
+brew install cmake swig sdl2
+pip install --upgrade pip setuptools wheel
+pip install -r requirements.txt
+```
+
+### 8.2 MPS 不稳定或报错
+
+强制 CPU：
+
+```bash
+python -m src.train_ppo --device cpu
+```
+
+说明：脚本会先做一次 MPS 张量 sanity check，失败自动回退 CPU。
+
+### 8.3 视频写入失败（ffmpeg/codec）
+
+1) 安装系统 ffmpeg：
+```bash
+brew install ffmpeg
+```
+
+2) 已降级保存帧序列时，手动转码：
+```bash
+ffmpeg -framerate 30 -i artifacts/videos/<name>_frames/frame_%06d.png -c:v libx264 -pix_fmt yuv420p artifacts/videos/<name>.mp4
+```
+
+### 8.4 图形窗口相关报错
+
+本工程默认 `rgb_array` 录制，不依赖 GUI 窗口。
+若仍遇到 SDL 问题，可显式设置：
+
+```bash
+export SDL_VIDEODRIVER=dummy
+python -m src.record_video --duration-sec 10
+```
+
+### 8.5 `size mismatch for action_net`（加载旧模型时报错）
+
+典型原因：旧 checkpoint 的动作空间与当前配置不同（如 `right_only`=5 动作，`simple`=7 动作）。
+
+可选修复：
+
+1) 保持和 checkpoint 一致的动作空间：
+```bash
+python -m src.train_ppo --init-model-path /Users/roog/Work/FNT/SpMr/mario-rl-mvp/artifacts/models/ppo_SuperMarioBros-1-1-v0_20260212_164717/ppo_mario_ckpt_150000_steps.zip --movement right_only
+```
+
+2) 若你确实要切动作空间，用部分初始化（跳过不兼容动作头）：
+```bash
+python -m src.train_ppo --init-model-path artifacts/models/latest_model.zip --movement simple --allow-partial-init
+```
+
+3) 或者直接不加载旧模型，从头训练新动作空间。
+
+## 9. 最小 smoke test（按顺序执行）
+
+```bash
+cd /Users/roog/Work/FNT/SpMr/mario-rl-mvp
+source .venv/bin/activate
+
+# 1) 训练 1e4 steps，并至少写出 checkpoint + final model
+python -m src.train_ppo \
+  --total-timesteps 10000 \
+  --save-freq 2000 \
+  --n-envs 1 \
+  --device cpu \
+  --movement right_only
+
+# 2) 快速评估
+python -m src.eval --episodes 2 --max-steps 2000
+
+# 3) 录制 10 秒视频
+python -m src.record_video --duration-sec 10 --fps 30
+```
+
+验收标准：
+- `artifacts/models/` 下有 `.zip` 模型
+- `artifacts/logs/` 下有 TensorBoard event 文件
+- `artifacts/videos/` 下有 `.mp4`（或失败时有 `_frames/` 帧序列）
--- a/mario-rl-mvp/requirements.txt
+++ b/mario-rl-mvp/requirements.txt
@@ -0,0 +1,12 @@
+torch==2.5.1
+stable-baselines3==2.3.2
+gym==0.26.2
+gymnasium==0.29.1
+shimmy==1.3.0
+gym-super-mario-bros==7.4.0
+nes-py==8.2.1
+opencv-python==4.10.0.84
+numpy==1.26.4
+tensorboard==2.18.0
+imageio==2.36.1
+imageio-ffmpeg==0.5.1
--- a/mario-rl-mvp/src/init.py
+++ b/mario-rl-mvp/src/init.py
@@ -0,0 +1 @@
+"""Mario RL MVP package."""
--- a/mario-rl-mvp/src/env.py
+++ b/mario-rl-mvp/src/env.py
@@ -0,0 +1,337 @@
+from __future__ import annotations
+
+from collections import deque
+from typing import Any, Callable, Deque, Dict, Optional, Tuple
+
+import cv2
+import gym
+import gym_super_mario_bros
+import numpy as np
+from gym.spaces import Box
+from gym_super_mario_bros.actions import RIGHT_ONLY, SIMPLE_MOVEMENT
+from nes_py.wrappers import JoypadSpace
+
+
+def reset_compat(env: gym.Env, seed: Optional[int] = None) -> Tuple[np.ndarray, Dict[str, Any]]:
+    try:
+        result = env.reset(seed=seed)
+    except TypeError:
+        result = env.reset()
+        if seed is not None and hasattr(env, "seed"):
+            env.seed(seed)
+
+    if isinstance(result, tuple) and len(result) == 2:
+        obs, info = result
+        return obs, info
+    return result, {}
+
+
+def step_compat(env: gym.Env, action: Any) -> Tuple[np.ndarray, float, bool, bool, Dict[str, Any]]:
+    result = env.step(action)
+    if isinstance(result, tuple) and len(result) == 5:
+        obs, reward, terminated, truncated, info = result
+        return obs, float(reward), bool(terminated), bool(truncated), info
+    if isinstance(result, tuple) and len(result) == 4:
+        obs, reward, done, info = result
+        return obs, float(reward), bool(done), False, info
+    raise RuntimeError(f"Unexpected step return format: {type(result)} / {result}")
+
+
+class SkipFrame(gym.Wrapper):
+    def __init__(self, env: gym.Env, skip: int = 4):
+        super().__init__(env)
+        self._skip = skip
+
+    def step(self, action: Any):
+        total_reward = 0.0
+        terminated = False
+        truncated = False
+        info: Dict[str, Any] = {}
+        obs = None
+        for _ in range(self._skip):
+            obs, reward, terminated, truncated, info = step_compat(self.env, action)
+            total_reward += reward
+            if terminated or truncated:
+                break
+        return obs, total_reward, terminated, truncated, info
+
+
+class PreprocessFrame(gym.ObservationWrapper):
+    """Convert RGB frame to grayscale 84x84 uint8."""
+
+    def __init__(self, env: gym.Env, width: int = 84, height: int = 84):
+        super().__init__(env)
+        self.width = width
+        self.height = height
+        self.observation_space = Box(low=0, high=255, shape=(height, width, 1), dtype=np.uint8)
+
+    def observation(self, observation: np.ndarray) -> np.ndarray:
+        if observation.ndim == 3 and observation.shape[2] == 3:
+            gray = cv2.cvtColor(observation, cv2.COLOR_RGB2GRAY)
+        elif observation.ndim == 2:
+            gray = observation
+        else:
+            gray = np.squeeze(observation)
+        resized = cv2.resize(gray, (self.width, self.height), interpolation=cv2.INTER_AREA)
+        return resized[:, :, None].astype(np.uint8)
+
+
+class ChannelLastFrameStack(gym.Wrapper):
+    """Stack frames on the channel axis: (H, W, C*num_stack)."""
+
+    def __init__(self, env: gym.Env, num_stack: int = 4):
+        super().__init__(env)
+        self.num_stack = num_stack
+        self.frames: Deque[np.ndarray] = deque(maxlen=num_stack)
+
+        obs_space = env.observation_space
+        assert isinstance(obs_space, Box), "Frame stack requires Box observation space."
+        h, w, c = obs_space.shape
+        self.observation_space = Box(
+            low=0,
+            high=255,
+            shape=(h, w, c * num_stack),
+            dtype=np.uint8,
+        )
+
+    def reset(self, *, seed: Optional[int] = None, options: Optional[dict] = None):
+        del options
+        obs, info = reset_compat(self.env, seed=seed)
+        self.frames.clear()
+        for _ in range(self.num_stack):
+            self.frames.append(obs)
+        return self._get_observation(), info
+
+    def step(self, action: Any):
+        obs, reward, terminated, truncated, info = step_compat(self.env, action)
+        self.frames.append(obs)
+        return self._get_observation(), reward, terminated, truncated, info
+
+    def _get_observation(self) -> np.ndarray:
+        assert len(self.frames) == self.num_stack
+        return np.concatenate(list(self.frames), axis=2)
+
+
+class TransposeObservation(gym.ObservationWrapper):
+    """Convert observation from HWC to CHW for CNN policy."""
+
+    def __init__(self, env: gym.Env):
+        super().__init__(env)
+        obs_space = env.observation_space
+        assert isinstance(obs_space, Box), "TransposeObservation requires Box observation space."
+        h, w, c = obs_space.shape
+        self.observation_space = Box(low=0, high=255, shape=(c, h, w), dtype=obs_space.dtype)
+
+    def observation(self, observation: np.ndarray) -> np.ndarray:
+        return np.transpose(observation, (2, 0, 1)).astype(np.uint8)
+
+
+class ClipRewardEnv(gym.RewardWrapper):
+    def reward(self, reward):
+        return float(np.sign(reward))
+
+
+class ProgressRewardEnv(gym.Wrapper):
+    """Reward shaping focused on moving right and avoiding local traps."""
+
+    def __init__(
+        self,
+        env: gym.Env,
+        progress_scale: float = 0.02,
+        death_penalty: float = -50.0,
+        flag_bonus: float = 100.0,
+        stall_penalty: float = 0.05,
+        stall_steps: int = 40,
+        backward_penalty_scale: float = 0.01,
+        milestone_interval: int = 32,
+        milestone_bonus: float = 1.0,
+        no_progress_terminate_steps: int = 120,
+        no_progress_terminate_penalty: float = 20.0,
+    ):
+        super().__init__(env)
+        self.progress_scale = progress_scale
+        self.death_penalty = death_penalty
+        self.flag_bonus = flag_bonus
+        self.stall_penalty = stall_penalty
+        self.stall_steps = stall_steps
+        self.backward_penalty_scale = backward_penalty_scale
+        self.milestone_interval = milestone_interval
+        self.milestone_bonus = milestone_bonus
+        self.no_progress_terminate_steps = no_progress_terminate_steps
+        self.no_progress_terminate_penalty = no_progress_terminate_penalty
+        self._last_x_pos: Optional[float] = None
+        self._best_x_pos = 0.0
+        self._stall_count = 0
+        self._next_milestone_x = float(milestone_interval)
+
+    def reset(self, *, seed: Optional[int] = None, options: Optional[dict] = None):
+        del options
+        obs, info = reset_compat(self.env, seed=seed)
+        self._last_x_pos = float(info.get("x_pos", 0.0))
+        self._best_x_pos = self._last_x_pos
+        self._stall_count = 0
+        if self.milestone_interval > 0:
+            k = int(self._best_x_pos // self.milestone_interval) + 1
+            self._next_milestone_x = float(k * self.milestone_interval)
+        return obs, info
+
+    def step(self, action: Any):
+        obs, reward, terminated, truncated, info = step_compat(self.env, action)
+        x_pos = float(info.get("x_pos", 0.0))
+
+        if self._last_x_pos is None:
+            delta_x = 0.0
+        else:
+            delta_x = x_pos - self._last_x_pos
+        self._last_x_pos = x_pos
+
+        shaped_reward = float(reward)
+        if delta_x > 0:
+            shaped_reward += self.progress_scale * delta_x
+            self._stall_count = 0
+            if x_pos > self._best_x_pos:
+                self._best_x_pos = x_pos
+            if self.milestone_interval > 0 and self.milestone_bonus != 0.0:
+                while x_pos >= self._next_milestone_x:
+                    shaped_reward += self.milestone_bonus
+                    self._next_milestone_x += self.milestone_interval
+        else:
+            self._stall_count += 1
+            if delta_x < 0:
+                shaped_reward -= self.backward_penalty_scale * abs(delta_x)
+
+        if self._stall_count >= self.stall_steps:
+            shaped_reward -= self.stall_penalty
+
+        if (
+            self.no_progress_terminate_steps > 0
+            and not terminated
+            and not truncated
+            and self._stall_count >= self.no_progress_terminate_steps
+        ):
+            truncated = True
+            shaped_reward -= self.no_progress_terminate_penalty
+            info["terminated_by_stall"] = True
+
+        if terminated or truncated:
+            if bool(info.get("flag_get", False)):
+                shaped_reward += self.flag_bonus
+            elif terminated:
+                shaped_reward += self.death_penalty
+
+        return obs, shaped_reward, terminated, truncated, info
+
+
+def get_action_set(name: str):
+    name = name.lower().strip()
+    if name == "simple":
+        return SIMPLE_MOVEMENT
+    if name == "right_only":
+        return RIGHT_ONLY
+    raise ValueError(f"Unsupported movement='{name}'. Use one of: right_only, simple")
+
+
+def make_mario_env(
+    env_id: str = "SuperMarioBros-1-1-v0",
+    seed: int = 0,
+    movement: str = "right_only",
+    reward_mode: str = "raw",
+    clip_reward: bool = False,
+    frame_skip: int = 4,
+    render_mode: Optional[str] = None,
+    progress_scale: float = 0.02,
+    death_penalty: float = -50.0,
+    flag_bonus: float = 100.0,
+    stall_penalty: float = 0.05,
+    stall_steps: int = 40,
+    backward_penalty_scale: float = 0.01,
+    milestone_interval: int = 32,
+    milestone_bonus: float = 1.0,
+    no_progress_terminate_steps: int = 120,
+    no_progress_terminate_penalty: float = 20.0,
+) -> gym.Env:
+    kwargs: Dict[str, Any] = {}
+    if render_mode is not None:
+        kwargs["render_mode"] = render_mode
+
+    try:
+        env = gym_super_mario_bros.make(env_id, apply_api_compatibility=True, **kwargs)
+    except TypeError:
+        env = gym_super_mario_bros.make(env_id, **kwargs)
+
+    env = JoypadSpace(env, get_action_set(movement))
+    env = SkipFrame(env, skip=frame_skip)
+
+    mode = reward_mode.lower().strip()
+    if clip_reward and mode == "raw":
+        mode = "clip"
+    if mode == "clip":
+        env = ClipRewardEnv(env)
+    elif mode == "progress":
+        env = ProgressRewardEnv(
+            env=env,
+            progress_scale=progress_scale,
+            death_penalty=death_penalty,
+            flag_bonus=flag_bonus,
+            stall_penalty=stall_penalty,
+            stall_steps=stall_steps,
+            backward_penalty_scale=backward_penalty_scale,
+            milestone_interval=milestone_interval,
+            milestone_bonus=milestone_bonus,
+            no_progress_terminate_steps=no_progress_terminate_steps,
+            no_progress_terminate_penalty=no_progress_terminate_penalty,
+        )
+    elif mode != "raw":
+        raise ValueError(f"Unsupported reward_mode='{reward_mode}'. Use one of: raw, clip, progress")
+
+    env = PreprocessFrame(env, width=84, height=84)
+    env = ChannelLastFrameStack(env, num_stack=4)
+    env = TransposeObservation(env)
+
+    # Seed once so each env subprocess/dummy env has deterministic startup.
+    reset_compat(env, seed=seed)
+    if hasattr(env.action_space, "seed"):
+        env.action_space.seed(seed)
+    return env
+
+
+def make_env_fn(
+    env_id: str,
+    seed: int,
+    movement: str,
+    reward_mode: str,
+    clip_reward: bool,
+    frame_skip: int,
+    progress_scale: float,
+    death_penalty: float,
+    flag_bonus: float,
+    stall_penalty: float,
+    stall_steps: int,
+    backward_penalty_scale: float,
+    milestone_interval: int,
+    milestone_bonus: float,
+    no_progress_terminate_steps: int,
+    no_progress_terminate_penalty: float,
+) -> Callable[[], gym.Env]:
+    def _thunk() -> gym.Env:
+        return make_mario_env(
+            env_id=env_id,
+            seed=seed,
+            movement=movement,
+            reward_mode=reward_mode,
+            clip_reward=clip_reward,
+            frame_skip=frame_skip,
+            render_mode=None,
+            progress_scale=progress_scale,
+            death_penalty=death_penalty,
+            flag_bonus=flag_bonus,
+            stall_penalty=stall_penalty,
+            stall_steps=stall_steps,
+            backward_penalty_scale=backward_penalty_scale,
+            milestone_interval=milestone_interval,
+            milestone_bonus=milestone_bonus,
+            no_progress_terminate_steps=no_progress_terminate_steps,
+            no_progress_terminate_penalty=no_progress_terminate_penalty,
+        )
+
+    return _thunk
--- a/mario-rl-mvp/src/eval.py
+++ b/mario-rl-mvp/src/eval.py
@@ -0,0 +1,162 @@
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+from statistics import mean
+
+import numpy as np
+from stable_baselines3 import PPO
+
+from src.env import get_action_set, make_mario_env, reset_compat, step_compat
+from src.utils import ensure_artifact_paths, latest_model_path, seed_everything
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Evaluate trained Mario PPO agent.")
+    parser.add_argument("--model-path", type=str, default="", help="Path to .zip model. If empty, use latest model.")
+    parser.add_argument("--env-id", type=str, default="SuperMarioBros-1-1-v0")
+    parser.add_argument("--movement", type=str, default="auto", choices=["auto", "right_only", "simple"])
+    parser.add_argument("--episodes", type=int, default=5)
+    parser.add_argument("--max-steps", type=int, default=5_000)
+    parser.add_argument("--seed", type=int, default=42)
+    parser.add_argument("--frame-skip", type=int, default=4)
+    parser.add_argument("--reward-mode", type=str, default="raw", choices=["raw", "clip", "progress"])
+    parser.add_argument("--progress-scale", type=float, default=0.02)
+    parser.add_argument("--death-penalty", type=float, default=-50.0)
+    parser.add_argument("--flag-bonus", type=float, default=100.0)
+    parser.add_argument("--stall-penalty", type=float, default=0.05)
+    parser.add_argument("--stall-steps", type=int, default=40)
+    parser.add_argument("--backward-penalty-scale", type=float, default=0.01)
+    parser.add_argument("--milestone-interval", type=int, default=32)
+    parser.add_argument("--milestone-bonus", type=float, default=1.0)
+    parser.add_argument("--no-progress-terminate-steps", type=int, default=120)
+    parser.add_argument("--no-progress-terminate-penalty", type=float, default=20.0)
+    parser.add_argument("--clip-reward", action="store_true")
+    parser.add_argument("--stochastic", action="store_true", help="Use stochastic policy (deterministic=False).")
+    return parser.parse_args()
+
+
+def _action_count_for_movement(movement: str) -> int:
+    return len(get_action_set(movement))
+
+
+def _model_action_count(model: PPO):
+    if hasattr(model, "action_space") and hasattr(model.action_space, "n"):
+        return int(model.action_space.n)
+    action_net = getattr(getattr(model, "policy", None), "action_net", None)
+    out_features = getattr(action_net, "out_features", None)
+    if out_features is not None:
+        return int(out_features)
+    return None
+
+
+def resolve_movement(movement_arg: str, model: PPO) -> str:
+    model_action_n = _model_action_count(model)
+
+    if movement_arg == "auto":
+        if model_action_n is not None:
+            for candidate in ("right_only", "simple"):
+                if _action_count_for_movement(candidate) == model_action_n:
+                    return candidate
+        return "right_only"
+
+    if model_action_n is not None:
+        expected_n = _action_count_for_movement(movement_arg)
+        if expected_n != model_action_n:
+            raise ValueError(
+                f"movement='{movement_arg}' has {expected_n} actions, but model expects {model_action_n}. "
+                "Use --movement auto or pass the matching movement."
+            )
+    return movement_arg
+
+
+def resolve_model_path(user_path: str) -> Path:
+    if user_path:
+        p = Path(user_path).expanduser().resolve()
+        if not p.exists():
+            raise FileNotFoundError(f"Model not found: {p}")
+        return p
+
+    paths = ensure_artifact_paths()
+    latest = latest_model_path(paths.models)
+    if latest is None:
+        raise FileNotFoundError("No model found under artifacts/models. Please run training first.")
+    return latest
+
+
+def main() -> None:
+    args = parse_args()
+    seed_everything(args.seed)
+    reward_mode = "clip" if args.clip_reward and args.reward_mode == "raw" else args.reward_mode
+
+    model_path = resolve_model_path(args.model_path)
+    print(f"[eval] model={model_path}")
+    model = PPO.load(str(model_path))
+    movement = resolve_movement(args.movement, model)
+    print(f"[eval] movement={movement} reward_mode={reward_mode}")
+
+    env = make_mario_env(
+        env_id=args.env_id,
+        seed=args.seed,
+        movement=movement,
+        reward_mode=reward_mode,
+        clip_reward=args.clip_reward,
+        frame_skip=args.frame_skip,
+        render_mode=None,
+        progress_scale=args.progress_scale,
+        death_penalty=args.death_penalty,
+        flag_bonus=args.flag_bonus,
+        stall_penalty=args.stall_penalty,
+        stall_steps=args.stall_steps,
+        backward_penalty_scale=args.backward_penalty_scale,
+        milestone_interval=args.milestone_interval,
+        milestone_bonus=args.milestone_bonus,
+        no_progress_terminate_steps=args.no_progress_terminate_steps,
+        no_progress_terminate_penalty=args.no_progress_terminate_penalty,
+    )
+
+    rewards = []
+    max_x_positions = []
+    clear_flags = []
+
+    for ep in range(1, args.episodes + 1):
+        obs, info = reset_compat(env, seed=args.seed + ep)
+        done = False
+        ep_reward = 0.0
+        ep_max_x = float(info.get("x_pos", 0.0))
+        flag_get = False
+        step_count = 0
+
+        while not done and step_count < args.max_steps:
+            action, _ = model.predict(obs, deterministic=not args.stochastic)
+            if isinstance(action, np.ndarray):
+                action = int(action.item())
+            obs, reward, terminated, truncated, info = step_compat(env, action)
+            ep_reward += float(reward)
+            ep_max_x = max(ep_max_x, float(info.get("x_pos", 0.0)))
+            flag_get = flag_get or bool(info.get("flag_get", False))
+            done = terminated or truncated
+            step_count += 1
+
+        rewards.append(ep_reward)
+        max_x_positions.append(ep_max_x)
+        clear_flags.append(1.0 if flag_get else 0.0)
+        print(
+            f"[episode {ep}] reward={ep_reward:.2f} max_x={ep_max_x:.1f} "
+            f"clear={flag_get} steps={step_count}"
+        )
+
+    summary = {
+        "episodes": args.episodes,
+        "avg_reward": mean(rewards) if rewards else 0.0,
+        "avg_max_x_pos": mean(max_x_positions) if max_x_positions else 0.0,
+        "clear_rate": mean(clear_flags) if clear_flags else 0.0,
+    }
+    print("[summary]", json.dumps(summary, ensure_ascii=False, indent=2))
+
+    env.close()
+
+
+if __name__ == "__main__":
+    main()
--- a/mario-rl-mvp/src/record_video.py
+++ b/mario-rl-mvp/src/record_video.py
@@ -0,0 +1,200 @@
+from __future__ import annotations
+
+import argparse
+from datetime import datetime
+from pathlib import Path
+
+import imageio.v2 as imageio
+import numpy as np
+from stable_baselines3 import PPO
+
+from src.env import get_action_set, make_mario_env, reset_compat, step_compat
+from src.utils import ensure_artifact_paths, latest_model_path, seed_everything
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Record Mario agent rollout to mp4 (headless).")
+    parser.add_argument("--model-path", type=str, default="", help="Path to .zip model. Empty means latest model.")
+    parser.add_argument("--env-id", type=str, default="SuperMarioBros-1-1-v0")
+    parser.add_argument("--movement", type=str, default="auto", choices=["auto", "right_only", "simple"])
+    parser.add_argument("--seed", type=int, default=42)
+    parser.add_argument("--frame-skip", type=int, default=4)
+    parser.add_argument("--reward-mode", type=str, default="raw", choices=["raw", "clip", "progress"])
+    parser.add_argument("--progress-scale", type=float, default=0.02)
+    parser.add_argument("--death-penalty", type=float, default=-50.0)
+    parser.add_argument("--flag-bonus", type=float, default=100.0)
+    parser.add_argument("--stall-penalty", type=float, default=0.05)
+    parser.add_argument("--stall-steps", type=int, default=40)
+    parser.add_argument("--backward-penalty-scale", type=float, default=0.01)
+    parser.add_argument("--milestone-interval", type=int, default=32)
+    parser.add_argument("--milestone-bonus", type=float, default=1.0)
+    parser.add_argument("--no-progress-terminate-steps", type=int, default=120)
+    parser.add_argument("--no-progress-terminate-penalty", type=float, default=20.0)
+    parser.add_argument("--clip-reward", action="store_true")
+    parser.add_argument("--fps", type=int, default=30)
+    parser.add_argument("--duration-sec", type=int, default=10)
+    parser.add_argument("--max-steps", type=int, default=5_000)
+    parser.add_argument("--stochastic", action="store_true")
+    parser.add_argument("--output", type=str, default="", help="Output mp4 path.")
+    return parser.parse_args()
+
+
+def _action_count_for_movement(movement: str) -> int:
+    return len(get_action_set(movement))
+
+
+def _model_action_count(model: PPO):
+    if hasattr(model, "action_space") and hasattr(model.action_space, "n"):
+        return int(model.action_space.n)
+    action_net = getattr(getattr(model, "policy", None), "action_net", None)
+    out_features = getattr(action_net, "out_features", None)
+    if out_features is not None:
+        return int(out_features)
+    return None
+
+
+def resolve_movement(movement_arg: str, model: PPO) -> str:
+    model_action_n = _model_action_count(model)
+
+    if movement_arg == "auto":
+        if model_action_n is not None:
+            for candidate in ("right_only", "simple"):
+                if _action_count_for_movement(candidate) == model_action_n:
+                    return candidate
+        return "right_only"
+
+    if model_action_n is not None:
+        expected_n = _action_count_for_movement(movement_arg)
+        if expected_n != model_action_n:
+            raise ValueError(
+                f"movement='{movement_arg}' has {expected_n} actions, but model expects {model_action_n}. "
+                "Use --movement auto or pass the matching movement."
+            )
+    return movement_arg
+
+
+def resolve_model(user_path: str) -> Path:
+    if user_path:
+        p = Path(user_path).expanduser().resolve()
+        if not p.exists():
+            raise FileNotFoundError(f"Model not found: {p}")
+        return p
+
+    paths = ensure_artifact_paths()
+    latest = latest_model_path(paths.models)
+    if latest is None:
+        raise FileNotFoundError("No model found under artifacts/models. Please run training first.")
+    return latest
+
+
+def resolve_output_path(user_output: str) -> Path:
+    if user_output:
+        return Path(user_output).expanduser().resolve()
+
+    paths = ensure_artifact_paths()
+    ts = datetime.now().strftime("%Y%m%d_%H%M%S")
+    return (paths.videos / f"mario_replay_{ts}.mp4").resolve()
+
+
+def save_frames_fallback(frames, output_path: Path) -> Path:
+    frame_dir = output_path.with_suffix("")
+    frame_dir.mkdir(parents=True, exist_ok=True)
+    for i, frame in enumerate(frames):
+        imageio.imwrite(frame_dir / f"frame_{i:06d}.png", frame)
+    return frame_dir
+
+
+def main() -> None:
+    args = parse_args()
+    seed_everything(args.seed)
+    reward_mode = "clip" if args.clip_reward and args.reward_mode == "raw" else args.reward_mode
+
+    model_path = resolve_model(args.model_path)
+    output_path = resolve_output_path(args.output)
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+
+    model = PPO.load(str(model_path))
+    movement = resolve_movement(args.movement, model)
+    print(f"[video] movement={movement} reward_mode={reward_mode}")
+    env = make_mario_env(
+        env_id=args.env_id,
+        seed=args.seed,
+        movement=movement,
+        reward_mode=reward_mode,
+        clip_reward=args.clip_reward,
+        frame_skip=args.frame_skip,
+        render_mode="rgb_array",
+        progress_scale=args.progress_scale,
+        death_penalty=args.death_penalty,
+        flag_bonus=args.flag_bonus,
+        stall_penalty=args.stall_penalty,
+        stall_steps=args.stall_steps,
+        backward_penalty_scale=args.backward_penalty_scale,
+        milestone_interval=args.milestone_interval,
+        milestone_bonus=args.milestone_bonus,
+        no_progress_terminate_steps=args.no_progress_terminate_steps,
+        no_progress_terminate_penalty=args.no_progress_terminate_penalty,
+    )
+
+    obs, _ = reset_compat(env, seed=args.seed)
+    frames = []
+
+    first_frame = env.render()
+    if first_frame is not None:
+        # nes-py may reuse the same frame buffer; copy to avoid aliasing all frames.
+        frames.append(first_frame.copy())
+
+    target_frames = max(1, args.fps * args.duration_sec)
+    done = False
+    step_count = 0
+    reward_sum = 0.0
+    episode_count = 1
+
+    while len(frames) < target_frames and step_count < args.max_steps:
+        action, _ = model.predict(obs, deterministic=not args.stochastic)
+        if isinstance(action, np.ndarray):
+            action = int(action.item())
+        obs, reward, terminated, truncated, info = step_compat(env, action)
+        reward_sum += float(reward)
+        step_count += 1
+        done = terminated or truncated
+
+        frame = env.render()
+        if frame is not None:
+            frames.append(frame.copy())
+
+        if done:
+            obs, _ = reset_compat(env, seed=args.seed + episode_count)
+            episode_count += 1
+            frame = env.render()
+            if frame is not None and len(frames) < target_frames:
+                frames.append(frame.copy())
+            done = False
+
+    if not frames:
+        raise RuntimeError("No frames captured. Check render_mode support and environment setup.")
+
+    try:
+        writer = imageio.get_writer(str(output_path), fps=args.fps, codec="libx264", quality=8)
+        for frame in frames:
+            writer.append_data(frame)
+        writer.close()
+        print(f"[video] Saved mp4: {output_path}")
+    except Exception as exc:
+        frame_dir = save_frames_fallback(frames, output_path)
+        print(f"[warn] mp4 write failed: {exc}")
+        print(f"[fallback] Saved frame sequence: {frame_dir}")
+        print(
+            "[fallback] Convert frames to mp4 with: "
+            f"ffmpeg -framerate {args.fps} -i {frame_dir}/frame_%06d.png -c:v libx264 -pix_fmt yuv420p {output_path}"
+        )
+
+    print(
+        f"[stats] frames={len(frames)} approx_sec={len(frames)/max(args.fps,1):.2f} "
+        f"steps={step_count} reward_sum={reward_sum:.2f} episodes={episode_count}"
+    )
+    env.close()
+
+
+if __name__ == "__main__":
+    main()
--- a/mario-rl-mvp/src/train_ppo.py
+++ b/mario-rl-mvp/src/train_ppo.py
@@ -0,0 +1,300 @@
+from __future__ import annotations
+
+import argparse
+import shutil
+from datetime import datetime
+from pathlib import Path
+
+from stable_baselines3 import PPO
+from stable_baselines3.common.callbacks import BaseCallback, CallbackList, CheckpointCallback
+from stable_baselines3.common.save_util import load_from_zip_file
+from stable_baselines3.common.vec_env import DummyVecEnv, VecMonitor
+from torch.utils.tensorboard import SummaryWriter
+
+from src.env import make_env_fn
+from src.utils import ensure_artifact_paths, resolve_torch_device, seed_everything, write_latest_pointer
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Train PPO agent on NES Super Mario Bros.")
+    parser.add_argument("--env-id", type=str, default="SuperMarioBros-1-1-v0")
+    parser.add_argument("--movement", type=str, default="right_only", choices=["right_only", "simple"])
+    parser.add_argument("--total-timesteps", type=int, default=1_000_000)
+    parser.add_argument("--seed", type=int, default=42)
+    parser.add_argument("--n-envs", type=int, default=4)
+    parser.add_argument("--frame-skip", type=int, default=4)
+    parser.add_argument("--reward-mode", type=str, default="raw", choices=["raw", "clip", "progress"])
+    parser.add_argument("--progress-scale", type=float, default=0.02)
+    parser.add_argument("--death-penalty", type=float, default=-50.0)
+    parser.add_argument("--flag-bonus", type=float, default=100.0)
+    parser.add_argument("--stall-penalty", type=float, default=0.05)
+    parser.add_argument("--stall-steps", type=int, default=40)
+    parser.add_argument("--backward-penalty-scale", type=float, default=0.01)
+    parser.add_argument("--milestone-interval", type=int, default=32)
+    parser.add_argument("--milestone-bonus", type=float, default=1.0)
+    parser.add_argument("--no-progress-terminate-steps", type=int, default=120)
+    parser.add_argument("--no-progress-terminate-penalty", type=float, default=20.0)
+
+    parser.add_argument("--learning-rate", type=float, default=2.5e-4)
+    parser.add_argument("--n-steps", type=int, default=128)
+    parser.add_argument("--batch-size", type=int, default=256)
+    parser.add_argument("--n-epochs", type=int, default=4)
+    parser.add_argument("--gamma", type=float, default=0.99)
+    parser.add_argument("--gae-lambda", type=float, default=0.95)
+    parser.add_argument("--ent-coef", type=float, default=0.01)
+    parser.add_argument("--clip-range", type=float, default=0.2)
+
+    parser.add_argument("--save-freq", type=int, default=50_000, help="Checkpoint frequency in env steps.")
+    parser.add_argument("--device", type=str, default="auto", choices=["auto", "cpu", "mps"])
+    parser.add_argument("--clip-reward", action="store_true", help="Enable reward clipping to {-1,0,1}.")
+    parser.add_argument("--run-name", type=str, default="", help="Optional custom run name.")
+    parser.add_argument(
+        "--init-model-path",
+        type=str,
+        default="",
+        help="Optional .zip model path to initialize weights before training.",
+    )
+    parser.add_argument(
+        "--allow-partial-init",
+        action="store_true",
+        help=(
+            "Allow partial init when checkpoint is from a different action space "
+            "(e.g. right_only -> simple). Compatible policy layers will be loaded, "
+            "incompatible layers (like action head) are skipped."
+        ),
+    )
+    parser.add_argument("--progress-bar", action="store_true", help="Enable SB3 progress bar (requires tqdm/rich).")
+    return parser.parse_args()
+
+
+def load_partial_policy_weights(model: PPO, init_model_path: Path, device: str) -> tuple[int, int]:
+    _, params, _ = load_from_zip_file(str(init_model_path), device=device)
+    if params is None or "policy" not in params:
+        raise RuntimeError(f"No policy parameters found in checkpoint: {init_model_path}")
+
+    src_state = params["policy"]
+    dst_state = model.policy.state_dict()
+
+    compatible_state = {}
+    skipped_count = 0
+    for key, value in src_state.items():
+        if key in dst_state and dst_state[key].shape == value.shape:
+            compatible_state[key] = value
+        else:
+            skipped_count += 1
+
+    if not compatible_state:
+        raise RuntimeError("No compatible policy tensors found for partial init.")
+
+    model.policy.load_state_dict(compatible_state, strict=False)
+    return len(compatible_state), skipped_count
+
+
+class EpisodeEndLoggingCallback(BaseCallback):
+    """Log per-episode terminal diagnostics to stdout and TensorBoard."""
+
+    REASON_TO_CODE = {"death": 0, "no_progress": 1, "timeout": 2, "clear": 3}
+
+    def __init__(self, n_envs: int, tb_log_dir: Path):
+        super().__init__(verbose=0)
+        self.n_envs = n_envs
+        self.tb_log_dir = tb_log_dir
+        self.tb_writer: SummaryWriter | None = None
+        self.episode_count = 0
+        self.episode_max_x_pos = [0.0 for _ in range(max(n_envs, 1))]
+
+    @staticmethod
+    def _resolve_done_reason(info: dict) -> str:
+        if bool(info.get("flag_get", False)):
+            return "clear"
+        if bool(info.get("terminated_by_stall", False)):
+            return "no_progress"
+        if bool(info.get("TimeLimit.truncated", False)):
+            return "timeout"
+        try:
+            if float(info.get("time", 1.0)) <= 0.0:
+                return "timeout"
+        except (TypeError, ValueError):
+            pass
+        return "death"
+
+    def _on_training_start(self) -> None:
+        self.tb_log_dir.mkdir(parents=True, exist_ok=True)
+        self.tb_writer = SummaryWriter(log_dir=str(self.tb_log_dir))
+
+    def _on_step(self) -> bool:
+        infos = self.locals.get("infos")
+        dones = self.locals.get("dones")
+        if infos is None or dones is None:
+            return True
+
+        for env_idx, (info, done) in enumerate(zip(infos, dones)):
+            if env_idx >= len(self.episode_max_x_pos):
+                continue
+            x_pos = float(info.get("x_pos", 0.0))
+            if x_pos > self.episode_max_x_pos[env_idx]:
+                self.episode_max_x_pos[env_idx] = x_pos
+
+            if not bool(done):
+                continue
+
+            self.episode_count += 1
+            max_x_pos = self.episode_max_x_pos[env_idx]
+            flag_get = 1.0 if bool(info.get("flag_get", False)) else 0.0
+            done_reason = self._resolve_done_reason(info)
+            done_reason_code = float(self.REASON_TO_CODE[done_reason])
+            episode_step = self.episode_count
+
+            self.logger.record_mean("rollout/episode_max_x_pos", max_x_pos)
+            self.logger.record_mean("rollout/flag_get", flag_get)
+            self.logger.record_mean(f"rollout/done_reason_{done_reason}", 1.0)
+
+            if self.tb_writer is not None:
+                self.tb_writer.add_scalar("episode_end/episode_max_x_pos", max_x_pos, episode_step)
+                self.tb_writer.add_scalar("episode_end/flag_get", flag_get, episode_step)
+                self.tb_writer.add_scalar("episode_end/done_reason_code", done_reason_code, episode_step)
+                self.tb_writer.add_scalar("episode_end/done_reason_death", 1.0 if done_reason == "death" else 0.0, episode_step)
+                self.tb_writer.add_scalar(
+                    "episode_end/done_reason_no_progress", 1.0 if done_reason == "no_progress" else 0.0, episode_step
+                )
+                self.tb_writer.add_scalar("episode_end/done_reason_timeout", 1.0 if done_reason == "timeout" else 0.0, episode_step)
+                self.tb_writer.add_scalar("episode_end/done_reason_clear", 1.0 if done_reason == "clear" else 0.0, episode_step)
+                self.tb_writer.flush()
+
+            print(
+                f"[episode_end] ep={episode_step} env={env_idx} reason={done_reason} "
+                f"max_x={max_x_pos:.1f} flag_get={bool(flag_get)}"
+            )
+            self.episode_max_x_pos[env_idx] = 0.0
+
+        return True
+
+    def _on_training_end(self) -> None:
+        if self.tb_writer is not None:
+            self.tb_writer.close()
+            self.tb_writer = None
+
+
+def main() -> None:
+    args = parse_args()
+    seed_everything(args.seed)
+    reward_mode = "clip" if args.clip_reward and args.reward_mode == "raw" else args.reward_mode
+
+    paths = ensure_artifact_paths()
+    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+    env_slug = args.env_id.replace("/", "_")
+    run_name = args.run_name.strip() or f"ppo_{env_slug}_{timestamp}"
+
+    run_model_dir = paths.models / run_name
+    run_log_dir = paths.logs / run_name
+    run_tb_dir = run_log_dir / "tb"
+    run_model_dir.mkdir(parents=True, exist_ok=True)
+    run_tb_dir.mkdir(parents=True, exist_ok=True)
+
+    device, device_msg = resolve_torch_device(args.device)
+    print(f"[device] {device} | {device_msg}")
+    print(f"[run] {run_name}")
+    print(
+        f"[config] env_id={args.env_id}, movement={args.movement}, reward_mode={reward_mode}, "
+        f"clip_reward={args.clip_reward}, n_envs={args.n_envs}"
+    )
+
+    env_fns = [
+        make_env_fn(
+            env_id=args.env_id,
+            seed=args.seed + i,
+            movement=args.movement,
+            reward_mode=reward_mode,
+            clip_reward=args.clip_reward,
+            frame_skip=args.frame_skip,
+            progress_scale=args.progress_scale,
+            death_penalty=args.death_penalty,
+            flag_bonus=args.flag_bonus,
+            stall_penalty=args.stall_penalty,
+            stall_steps=args.stall_steps,
+            backward_penalty_scale=args.backward_penalty_scale,
+            milestone_interval=args.milestone_interval,
+            milestone_bonus=args.milestone_bonus,
+            no_progress_terminate_steps=args.no_progress_terminate_steps,
+            no_progress_terminate_penalty=args.no_progress_terminate_penalty,
+        )
+        for i in range(args.n_envs)
+    ]
+    vec_env = DummyVecEnv(env_fns)
+    vec_env = VecMonitor(vec_env, filename=str(run_log_dir / "monitor.csv"))
+
+    model = PPO(
+        policy="CnnPolicy",
+        env=vec_env,
+        learning_rate=args.learning_rate,
+        n_steps=args.n_steps,
+        batch_size=args.batch_size,
+        n_epochs=args.n_epochs,
+        gamma=args.gamma,
+        gae_lambda=args.gae_lambda,
+        ent_coef=args.ent_coef,
+        clip_range=args.clip_range,
+        tensorboard_log=str(run_tb_dir),
+        seed=args.seed,
+        verbose=1,
+        device=device,
+    )
+
+    if args.init_model_path.strip():
+        init_model_path = Path(args.init_model_path).expanduser().resolve()
+        if not init_model_path.exists():
+            raise FileNotFoundError(f"Init model not found: {init_model_path}")
+        try:
+            model.set_parameters(str(init_model_path), exact_match=False, device=device)
+            print(f"[init] Loaded initial weights from: {init_model_path}")
+        except RuntimeError as exc:
+            if not args.allow_partial_init:
+                raise RuntimeError(
+                    f"{exc}\n"
+                    "Hint: checkpoint and current env may use different action spaces. "
+                    "Try one of:\n"
+                    "1) keep the same --movement as the checkpoint;\n"
+                    "2) remove --init-model-path and train from scratch;\n"
+                    "3) add --allow-partial-init to load compatible layers only."
+                ) from exc
+            loaded_count, skipped_count = load_partial_policy_weights(model, init_model_path, device=device)
+            print(
+                f"[init] Partial init from {init_model_path}: "
+                f"loaded_tensors={loaded_count}, skipped_tensors={skipped_count}"
+            )
+
+    callback = CheckpointCallback(
+        save_freq=max(args.save_freq // max(args.n_envs, 1), 1),
+        save_path=str(run_model_dir),
+        name_prefix="ppo_mario_ckpt",
+        save_replay_buffer=False,
+        save_vecnormalize=False,
+    )
+    episode_end_logging_callback = EpisodeEndLoggingCallback(
+        n_envs=args.n_envs,
+        tb_log_dir=run_tb_dir / "episode_end",
+    )
+
+    try:
+        model.learn(
+            total_timesteps=args.total_timesteps,
+            callback=CallbackList([callback, episode_end_logging_callback]),
+            tb_log_name="ppo",
+            progress_bar=args.progress_bar,
+        )
+    finally:
+        vec_env.close()
+
+    final_model = run_model_dir / "final_model"
+    model.save(str(final_model))
+
+    latest_model = paths.models / "latest_model.zip"
+    shutil.copy2(str(final_model) + ".zip", latest_model)
+    pointer = write_latest_pointer(paths.models, latest_model)
+
+    print(f"[done] Final model: {final_model}.zip")
+    print(f"[done] Latest model pointer: {pointer}")
+
+
+if __name__ == "__main__":
+    main()
--- a/mario-rl-mvp/src/utils.py
+++ b/mario-rl-mvp/src/utils.py
@@ -0,0 +1,75 @@
+from __future__ import annotations
+
+import random
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Optional, Tuple
+
+import numpy as np
+import torch
+
+
+@dataclass(frozen=True)
+class ArtifactPaths:
+    root: Path
+    models: Path
+    videos: Path
+    logs: Path
+
+
+def project_root() -> Path:
+    return Path(__file__).resolve().parents[1]
+
+
+def ensure_artifact_paths(root: Optional[Path] = None) -> ArtifactPaths:
+    root = root or project_root()
+    artifacts = root / "artifacts"
+    models = artifacts / "models"
+    videos = artifacts / "videos"
+    logs = artifacts / "logs"
+    for p in (artifacts, models, videos, logs):
+        p.mkdir(parents=True, exist_ok=True)
+    return ArtifactPaths(root=artifacts, models=models, videos=videos, logs=logs)
+
+
+def seed_everything(seed: int) -> None:
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(seed)
+
+
+def resolve_torch_device(requested: str = "auto") -> Tuple[str, str]:
+    requested = requested.lower().strip()
+    if requested == "cpu":
+        return "cpu", "User requested CPU."
+
+    if requested not in {"auto", "mps"}:
+        return "cpu", f"Unknown device '{requested}', fallback to CPU."
+
+    mps_available = hasattr(torch.backends, "mps") and torch.backends.mps.is_available()
+    if not mps_available:
+        if requested == "mps":
+            return "cpu", "MPS requested but unavailable, fallback to CPU."
+        return "cpu", "MPS unavailable, using CPU."
+
+    try:
+        x = torch.ones(8, device="mps")
+        _ = (x * 2).cpu().numpy()
+        return "mps", "MPS is available and passed a quick tensor sanity check."
+    except Exception as exc:  # pragma: no cover - hardware dependent
+        return "cpu", f"MPS check failed ({exc}), fallback to CPU."
+
+
+def latest_model_path(models_dir: Path) -> Optional[Path]:
+    candidates = [p for p in models_dir.rglob("*.zip") if p.is_file()]
+    if not candidates:
+        return None
+    return max(candidates, key=lambda p: p.stat().st_mtime)
+
+
+def write_latest_pointer(models_dir: Path, model_path: Path) -> Path:
+    pointer = models_dir / "LATEST_MODEL.txt"
+    pointer.write_text(str(model_path.resolve()) + "\n", encoding="utf-8")
+    return pointer