Skip to main content
Back

AgenticRL Lets GPT Watch Drones Fail and Rewrite Their Own Training

AgenticRL uses a multimodal GPT agent to write, test, and rewrite its own reward functions for drone navigation — hitting 91% real-world success with zero manual reward engineering

Paper, Full Access
  • AgenticRL is a framework where a multimodal GPT agent writes, critiques, and rewrites RL reward functions for drone navigation — no human reward engineering needed.
  • A "diagnosis packet" of training curves, trajectory heatmaps, and failure stats is fed back to the GPT agent, which identifies failure modes and revises the reward code in a closed loop.
  • The closed-loop refinement improves policy behavior by 71% over the initial GPT-generated reward across five navigation tasks.
  • Sim-to-real transfer achieves a 91% real-world success rate and 94% sim-to-real accuracy on a physical drone.
  • At deployment, the same multimodal agent reads a camera image and selects the correct pre-trained policy automatically — no manual switching.
  • No code or models are publicly released; the framework requires API access to a multimodal GPT model and full PPO retraining between each refinement iteration. Read the paper.

Training a drone to fly through a gate, dodge obstacles, or land precisely after crossing a wall barrier sounds like a job for a robotics PhD and weeks of reward function tuning. AgenticRL wants to make that a job for a language model. Researchers have built a closed-loop framework where a multimodal GPT agent writes the reward function, trains a policy with it, watches the drone fail, diagnoses why, and rewrites the reward , all without a human in the loop.

The reward engineering bottleneck

Reinforcement learning for robotics has a well-known dirty secret: the algorithm itself is the easy part. The quality of reward functions is what drives success, since well-designed reward signals are essential for guiding the agent's learning process toward desired behaviors. In practice, this means a roboticist manually crafts a reward formula, runs training for hours, watches the drone crash in a new and creative way, tweaks a coefficient, and repeats. Designing a high-performing reward function remains a challenging task that often requires substantial manual input.

The community has been chipping away at this for a while. NVIDIA's Eureka showed that LLM-generated reward programs can outperform expert human-written ones on more than 80% of tasks, leading to an average performance improvement of more than 50%. But most of these approaches still treat reward design as a one-shot or few-shot text generation problem , they generate candidates, pick the best, and stop. AgenticRL takes the loop further by adding a vision-aware critic that watches the trained policy actually behave and feeds that observation back into reward revision.

How the loop works

The framework has three distinct phases that chain together automatically:

  1. Reward generation: A multimodal GPT agent receives a natural language task description and a visual snapshot of the scene. It writes a task-specific reward function as executable code , no templates, no human-defined metrics.