RLinf/RPent
RPENT: EMBODIED AGENTS FRAMEWORK
Open framework for building embodied agents that evolve through recursive interaction with the physical world. Unifies perception, reasoning, memory, and execution into a single agentic system.
// KEY FEATURES
- Recursive architecture with memory, reflection, and self-evolution
- 92.63% success on LIBERO-PRO (741/800 tasks)
- Data flywheel: +24 pp on RoboTwin (48% → 72%)
- Support for real Franka robots and RoboCasa, RoboTwin simulators
// DEEP DIVE: HOW IT WORKS
RPent is a service-oriented framework where an agent consists of perception, reasoning, memory, and execution modules. The agent interacts with the physical world (simulator or real robot), collects trajectories, reflects on errors, and distills experience into memory to improve subsequent decisions. Planning can run via Claude Code, Codex, or a custom planner.
Key differentiator is memory-guided agentic composition: a frozen VLA model is steered by an agent with memory, raising success rates without retraining the base model. The data flywheel automatically gathers successful trajectories for SFT, yielding +24 pp on RoboTwin. The architecture is standardized and composable — easy to add new action primitives, simulators, or robots.
Requires GPU for VLA models, access to simulators (LIBERO, RoboCasa, RoboTwin) or real Franka robots (single/dual-arm). Non-reasoning mode cuts latency ~40%. The "Harness VLA" paper describes the method. Code in Python.
// USE CASES
Best for: Embodied AI researchers, RL engineers, robotics developers
- Develop embodied agents for long-horizon manipulation with memory
- Run data flywheel to auto-collect trajectories and fine-tune VLA
- Deploy agents on real Franka robots via standardized extensions
New repositories every 30 minutes
REDDYX AI scans GitHub 24/7 and ships the best AI/ML/Web3 projects to Telegram.
Join on Telegram