FARL: A Fully Automated Pipeline of RL Interface and Solver Design
We introduce FARL (Fully Automated Reinforcement Learning), a framework that automates RL design by jointly searching observation and reward programs, learning algorithms, and training settings.
ProofPaper ↗
Key points
- Existing reinforcement learning (RL) largely focuses on optimizing learning algorithms within a pre-defined task formulation, while applying RL to new tasks still relies heavily on human expertise to design the learning pipeline: what an agent observes, how it is rewarded, which algorithm it uses, and how training is configured.
- Automating these choices is challenging because they are tightly coupled.
- FARL uses training statistics and observed behavior to diagnose failures, propose changes, and test whether they improve learning.
- On Adroit Hand Relocate, FARL achieves 97.1% success at 5M training steps per policy, compared with 0% for Eureka and LiMEN and 1.0% for native Proximal Policy Optimization (PPO) under the same policy training budget.
Sources (1)
- [1]FARL: A Fully Automated Pipeline of RL Interface and Solver DesignarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 11, 03:22 PM
We introduce FARL (Fully Automated Reinforcement Learning), a framework that automates RL design by jointly searching observation and reward programs, learning algorithms, and training settings.
Existing reinforcement learning (RL) largely focuses on optimizing learning algorithms within a pre-defined task formulation, while applying RL to new tasks still relies heavily on human expertise to design the learning pipeline: what an agent observes, how it is rewarded, which algorithm it uses, and how training is configured.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 11, 2026RSIPO: Recursive Self-Improving Policy Optimization for Agentic Reinforcement Learning via Skill-Policy Co-Evolution
- Oct 10, 2026Self-Evolving Visual Agent via Margin Reward
- Oct 10, 2026Metric-Driven Reinforcement Learning for Object Hallucination Reduction in LVLM-Based Image Captioning
- Oct 10, 2026CFG-Policy: Unified Diffusion Policy Learning Framework via Classifier-free Guidance
- Oct 10, 2026From Video to Vision: Learning Generalizable Humanoid Carrying from Human Motion
- Oct 9, 2026Visualizing the World Model in an Agent's Value Function