Quickstart

Welcome to the RLinf Quickstart Guide. This section will walk you through launching RLinf for the first time. We present three concise examples to demonstrate the framework’s workflow and help you get started quickly.

• Installation: Two installation methods for RLinf are supported: using a Docker image or a custom user environment (see Installation).

• Embodied training: Training in the ManiSkill3 environment with the OpenVLA and OpenVLA-OFT models using the PPO algorithm (see Quickstart 1: PPO Training of VLAs on Maniskill3).

• Agentic training: Training on the boba dataset with the DeepSeek-R1-Distill-Qwen-1.5B model using the GRPO algorithm (see Quickstart 2: GRPO Training of LLMs on MATH).

• Distributed training: Multi-node training for embodied/agentic tasks (see Multi-node Training).

• Evaluation: Assessing model performance on embodied intelligence (see Evaluation Tutorial 1: Embodied VLA) and assessing model performance on long-chain-of-thought agentic tasks (see Evaluation Tutorial 2: Math Reasoning LLM).

SOTA RL Training Reproduction

RLinf provides end-to-end recipes that reproduce or match state-of-the-art (SOTA) RL results out of the box—users can directly run our configs and scripts to obtain published numbers without custom engineering.

For embodied tasks, RLinf reaches or matches SOTA success rates on benchmarks such as LIBERO, ManiSkill, RoboTwin, and more with OpenVLA, OpenVLA-OFT, π₀/π₀.₅, GR00T and other VLAs (see the Embodied Scenarios gallery and Supported RL Algorithms for algorithm details).

For agentic tasks (including math reasoning), RLinf achieves SOTA performance on AIME24/AIME25/GPQA-diamond benchmarks with DeepSeek-R1-Distill-Qwen models, and supports single-agent and multi-agent training tasks such as Search-R1 and Coding-Online-RL (see Agentic Scenarios).