RLinf/RLinf
RL INFRASTRUCTURE FOR EMBODIED AND AGENTIC AI
Scalable reinforcement learning infrastructure for robotic and agentic systems. Supports real robots, simulators, and VLA models with distributed rollout.
// KEY FEATURES
- Real robot support: Robbyant R2, LeKiwi, ARX X5, SO-101 with teleoperation
- Distributed rollout and training: 8 accelerator backends, PhyAI, NVIDIA Isaac integration
- VLA models and algorithms: OpenPI π₀/π₀.₅, RLT, STEAM, OPD, RTC, DiffusionART, Psi0
- Fault tolerance and CEL: hardware diagnostics, emergency stop, fleet collective learning
// DEEP DIVE: HOW IT WORKS
RLinf is a modular platform with rlinf.robotics core for physical robots and accelerated rollout backends. Cycle from Pico VR demonstrations to distributed PPO/GRPO on GPU. Fault-tolerance interface: diagnostics, auto-recovery, safe episode termination.
Unifies simulation (MuJoCo Warp, GenieSim, SIMPLE, RoboDojo), real robots (Robbyant R2, LeKiwi, ARX X5, SO-101), and models (OpenPI π₀/π₀.₅, Cosmos3, FastWAM, SD3, Wan2.2) in one stack. v0.4 added RLT, STEAM, OPD, RTC for real-world post-training and CEL — fleet collective learning without new data.
Requires GPU cluster (tested on H800) and compatible robots; setup includes calibration, bring-up, and Gym registration. Docs cover NVIDIA Isaac RL integration but assume RL and robotics expertise. Primary OS is Linux, Windows support not stated.
// USE CASES
Best for: For RL and robotics engineers to train policies on hardware and in simulation
- Train manipulation policy on dual-arm ARX X5 robot with Pico teleoperation
- Run distributed PPO for π₀.₅ in GenieSim with PhyAI rollout backend
- Deploy fault-tolerant rollout on Robbyant R2 fleet with CEL collective learning
New repositories every 30 minutes
REDDYX AI scans GitHub 24/7 and ships the best AI/ML/Web3 projects to Telegram.
Join on Telegram