RLinf/RPent

RPent: Agentic Infrastructure for the Physical World

What it solves

RPent addresses the challenge of making embodied agents (robots) more reliable and capable of handling long-horizon manipulation tasks. It solves the limitation where frozen Vision-Language-Action (VLA) models alone often struggle with complex tasks by wrapping them in an agentic framework that provides reasoning, memory, and self-evolution capabilities.

How it works

RPent is a recursive agent framework that treats various AI capabilities—perception, reasoning, memory, execution, and self-evolution—as composable services. It uses an "Agentic Planner" (such as Claude Code or Codex) to steer frozen VLA models, treating the VLAs as reliable manipulation primitives. The system operates through a loop of continuous interaction, reflection, and adaptation, allowing the agent to acquire new capabilities and evolve beyond its initial design through memory-guided interaction.

Who it’s for

  • Embodied intelligence researchers focusing on high success rates in long-horizon manipulation tasks.
  • Online and reinforcement learning researchers studying self-evolving agents and memory-distillation loops.
  • Robotics application developers who need a standardized, service-oriented architecture to deploy solutions on real hardware (e.g., Franka arms).
  • End users who want to run predefined robotic tasks out-of-the-box without needing deep machine learning expertise.

Highlights

  • Broad Compatibility: Supports multiple simulators (LIBERO-PRO, RoboCasa, RoboTwin, RoboDojo) and real-world hardware (Franka single and dual-arm).
  • Harness VLA: Specifically designed to steer frozen VLAs into reliable manipulation primitives.
  • Recursive Evolution: Enables agents to evolve through reflection and recursive interaction with the physical world.
  • Developer Tooling: Includes an interactive CLI for live steering and a Live Dashboard for streaming reasoning, camera views, and action timelines.

Related

  • Project
    AetherLabsAI/RSIAgent

    RSIAgent is a research‑grade, training‑free framework that lets LLMs recursively improve by exploring digital environments. Three agents (Curriculum, Actor, Verifier) run a broad‑then‑deep self‑improvement loop, storing distilled procedures in persistent memory that is frozen for downstream tasks. The repo includes full OSWorld‑V2 and Agents’ Last Exam (ALE) benchmark integrations, scripts for setup, validation, and batch execution, and documentation of the architecture and results.

  • Project
    YeQing17-2026/OmniAgent

    An open-source self-evolving agent framework that dynamically updates its skills, memory, and brain model through interaction while employing a multi-layer security scaffold.

  • Project
    Memento-Teams/Memento-Skills

    A skill-centric agent runtime that allows AI agents to design, execute, and persist reusable skills to improve their performance on complex tasks through a reflection-based learning loop.

  • Project
    AgentR1/Agent-R1

    A modular framework for Agentic Reinforcement Learning that trains multi-step LLM agents by treating each interaction turn as a step-level MDP transition.

  • Project
    RLinf/RLinf

    A scalable reinforcement learning infrastructure designed for Embodied and Agentic AI, providing a unified backbone for training VLA models and deploying them across simulators and real-world robots.