Summary

提出 AgenticRL,一个多模态 GPT agent 驱动的闭环 RL 框架:自动生成 reward、训练 PPO policy、诊断 failure modes、迭代精修 reward,最终在真实 UAV 上实现 91% success rate 和 94% sim-to-real accuracy。

Problem & Motivation

UAV RL 导航的核心痛点:reward function 设计依赖人工反复调参,费时且不保证 task success。自然语言+视觉场景能更直观描述 desired behavior,但转化为 dense numerical reward 仍是难题——语义正确的 reward 可能诱导 unsafe/incomplete behavior。现有方法(demonstration、preference、human feedback)需要额外监督或 task-specific data,且缺乏 multimodal scene grounding 和 real-world deployment 验证。

Method

框架核心:5 阶段闭环 pipeline:

  1. Multimodal Task Understanding:接收语言指令 l0、视觉场景 I、observation spec O
  2. Reward Generation:GPT agent 生成可执行 Python reward R0(包含 task progress rewards、safety penalties、terminal terms)
  3. Policy Training:在定制 UAV simulator 用 PPO 训练 policy(网络 512→512→256→128,entropy annealing 0.1→0.001,训练 25M-70M steps)
  4. Policy Diagnosis:评估 policy 行为,聚合 collision events、landing accuracy、gate traversal success 等为 diagnosis packet D
  5. Reward Refinement:GPT agent 将 diagnosis packet 转化为 refinement prompt,识别 failure modes,更新 reward → 闭环迭代

推理时:Multimodal Scenario Registry —— agent 根据真实图像+语言信息识别 active scenario,自动选择对应 trained policy。

与已有工作区别(vs. Text2Reward、Code as Reward、Agents Trainer):

  • 同时具备 Reward Code Generation ✓、Visual Input ✓、Closed-loop Auto Refinement ✓、Task Diagnosis ✓、Real-world UAV Platform ✓

Key Results

仿真性能:collective SSR 97%

Sim-to-Real Transfer

  • collective RSR 91%
  • collective S2R accuracy 94%
  • Gate Traversal / Circular Motion:接近完美
  • Trajectory Following:S2R 97%
  • Obstacle Avoidance:RSR 82%
  • Barrier Crossing:RSR 89%

Reward Refinement 效果

  • RRI (Reward Refinement Improvement) 71% —— 相比初始 reward,精修后 policy behavior 显著提升
  • 训练曲线:精修后 convergence 更清晰(gate traversal、obstacle landing、wall barrier landing 任务)

Ablation Study

MethodTrajectory FollowGate TraversalObstacle+LandBarrier+LandCircular Motion
Zero-shot reward0%80%63%0%55%
Few-shot reward48%74%88%72%98%
w/o analyzer50%61%60%52%62%
w/o vision conditioning87%95%65%48%76%
Full AgenticRL95%100%96%93%98%

关键发现:

  • Zero-shot reward 在某些任务直接失败(trajectory follow 0%、barrier 0%)
  • Few-shot 有提升但整体仍低于 full framework
  • w/o analyzer(无 failure analysis)跨任务表现差,说明闭环精修是关键
  • w/o vision conditioning 在 obstacle/barrier 任务明显下降,证明视觉 grounding 重要性

Strengths & Weaknesses

Strengths

  1. 闭环精修机制:不是一次性生成 reward,而是通过 behavior diagnosis + GPT agent feedback 实现迭代改进,71% RRI 证明设计有效
  2. Multimodal Grounding:同时利用语言指令和视觉场景生成 reward,ablation 显示视觉输入对 spatial-constrained tasks(obstacle/barrier)至关重要
  3. Real-world Validation:在物理 quadrotor 上验证,91% RSR + 94% S2R 是硬指标,不玩 simulation-only 游戏
  4. Task Diversity:5 个不同导航任务(gate traversal、obstacle avoidance、barrier crossing、trajectory following、motion generation),覆盖不同的 spatial/safety/termination 要求

Weaknesses

  1. 依赖 GPT Agent 能力:reward quality 和 consistency 受限于 underlying multimodal agent(文中提到不同模型可能表现不一致)
  2. 计算开销:迭代训练需要多个 RL training cycles,复杂任务耗时显著(25M-70M steps per cycle)
  3. 仍需手动设计组件:termination conditions、observation spaces、safety constraints、task-specific environment settings 仍需人工
  4. 实验场景局限:controlled environments,未在 highly dynamic/diverse real-world settings 评估
  5. Insight 深度不足:71% RRI 是 aggregate metric,具体每个 refinement epoch 改了什么、为什么有效缺乏细致分析

与研究方向关联

  • agentic-RL:直接相关 —— LLM agent 作为 reward designer + critic,实现 self-improving RL pipeline
  • navigation:UAV 导航是 embodied AI 重要子领域
  • VLM:多模态 grounding 用于 reward generation

Mind Map

mindmap
  root((AgenticRL))
    Problem
      UAV reward design 依赖人工
      语言/视觉描述难以转化为 numerical reward
      现有方法缺乏 multimodal grounding
    Method
      5-stage closed-loop pipeline
      Multimodal GPT agent as reward designer
      Diagnosis packet for failure analysis
      Iterative reward refinement
      Scenario registry for inference
    Results
      SSR 97%, RSR 91%, S2R 94%
      RRI 71% improvement
      Real UAV deployment
      5 navigation tasks validated

Notes

  • 与 Eureka (Ma et al., 2023) 的区别:Eureka 专注 text-to-reward + human verification,AgenticRL 引入 visual grounding + automated diagnosis + UAV real-world validation
  • 与 ReSkill (2606-01619) 的关联:两者都是 RL-in-the-loop,但 ReSkill 侧重 skill creation,AgenticRL 侧重 reward refinement
  • 有价值的问题:GPT agent 的 failure diagnosis 是否能泛化到其他 embodied tasks(manipulation、mobile manipulation)?vision conditioning 的具体贡献是什么——是帮助识别 scene entities 还是影响 reward structure?