游戏AI智能助手全解析：从传统脚本到AI大模型的技术演进

// 传统FSM伪代码
public class EnemyController : MonoBehaviour {
    enum State { Patrol, Chase, Attack }
    State currentState = State.Patrol;
    
    void Update() {
        switch(currentState) {
            case State.Patrol:
                Patrol();           // 按固定路径移动
                if(DetectPlayer()) currentState = State.Chase;
                break;
            case State.Chase:
                ChasePlayer();      // 直线追击
                if(WithinAttackRange()) currentState = State.Attack;
                if(!DetectPlayer()) currentState = State.Patrol;
                break;
            case State.Attack:
                AttackPlayer();     // 执行攻击动画
                if(!WithinAttackRange()) currentState = State.Chase;
                break;
        }
    }
}

 创建Conda虚拟环境
conda create -n mlagents python=3.10.12
conda activate mlagents

 克隆ML-Agents仓库
git clone --branch release_23 https://github.com/Unity-Technologies/ml-agents.git
cd ml-agents

 安装Python依赖包
pip install -e ./ml-agents-envs
pip install -e ./ml-agents

using UnityEngine;
using Unity.MLAgents;
using Unity.MLAgents.Sensors;
using Unity.MLAgents.Actuators;

/// <summary>
/// 示例：一个简单的RollerBall AI智能体
/// 智能体需要通过强化学习学会收集目标并避开障碍
/// </summary>
public class RollerAgent : Agent
{
    [SerializeField] private Transform target;    // 目标物体
    [SerializeField] private float forceMultiplier = 10f;
    private Rigidbody rBody;
    
    public override void Initialize()
    {
        rBody = GetComponent<Rigidbody>();
    }
    
    /// <summary>
    /// Step 1: 收集环境观测数据（状态空间）
    /// 智能体通过传感器获取位置信息
    /// </summary>
    public override void CollectObservations(VectorSensor sensor)
    {
        // 观测1-3：智能体自身位置
        sensor.AddObservation(transform.localPosition);
        // 观测4-6：目标物体位置
        sensor.AddObservation(target.localPosition);
        // 观测7-9：智能体速度
        sensor.AddObservation(rBody.velocity.x);
        sensor.AddObservation(rBody.velocity.z);
    }
    
    /// <summary>
    /// Step 2: 执行动作（动作空间）
    /// 智能体接收神经网络输出的动作指令
    /// </summary>
    public override void OnActionReceived(ActionBuffers actions)
    {
        // 动作空间：两个连续值，分别对应X轴和Z轴移动
        float moveX = actions.ContinuousActions[0];
        float moveZ = actions.ContinuousActions[1];
        
        Vector3 move = new Vector3(moveX, 0, moveZ)  forceMultiplier;
        rBody.AddForce(move);
        
        // Step 3: 惩罚时间成本，鼓励尽快完成任务
        AddReward(-1f / MaxStep);
    }
    
    /// <summary>
    /// Step 4: 定义奖励机制（核心学习信号）
    /// 触碰目标：+1.0分，并重置
    /// 掉落平台：-1.0分，并重置
    /// </summary>
    private void OnTriggerEnter(Collider other)
    {
        if (other.CompareTag("Target"))
        {
            AddReward(1.0f);           // 正面奖励
            target.localPosition = GetRandomPosition();
        }
        else if (other.CompareTag("Boundary"))
        {
            AddReward(-1.0f);          // 负面惩罚
            EndEpisode();              // 结束本轮训练
        }
    }
    
    public override void Heuristic(in ActionBuffers actionsOut)
    {
        // 人工控制模式（用于测试或录制演示数据）
        var continuousActions = actionsOut.ContinuousActions;
        continuousActions[0] = Input.GetAxis("Horizontal");
        continuousActions[1] = Input.GetAxis("Vertical");
    }
}

 config/rollerball_config.yaml
behaviors:
  RollerBall:
    trainer_type: ppo                     使用PPO算法
    hyperparameters:
      batch_size: 128                     每批训练样本数
      buffer_size: 2048                   经验回放缓冲区大小
      learning_rate: 0.0003               学习率
      beta: 0.005                         熵正则系数
      epsilon: 0.2                        PPO裁剪阈值
    network_settings:
      normalize: true                     归一化输入
      hidden_units: 128                   隐藏层神经元数
      num_layers: 2                       隐藏层层数
    max_steps: 500000                     最大训练步数
    time_horizon: 64                      时间窗口
    summary_freq: 10000                   日志输出频率

 在Unity中打开场景，然后执行
mlagents-learn config/rollerball_config.yaml --run-id=first_training