{"full_name":"RLinf/RLinf","owner":"RLinf","name":"RLinf","shard":"1","url":"https://github.com/RLinf/RLinf","description":"RLinf: Reinforcement Learning Infrastructure for Embodied and Agentic AI","homepage":"rlinf.org","topics":["reinforcement-learning","agentic-ai","rlinf","vla-rl","embodied-ai","rl-infra"],"tag_pages":["agentic-ai","reinforcement-learning"],"language":"Python","license":"Apache-2.0","avatar":"https://avatars.githubusercontent.com/u/226440105?v=4","created_at":"2025-08-14T03:52:19Z","pushed_at":"2026-11-15T14:44:57Z","metrics":{"stars":5654,"forks":845,"watchers":17,"issues":438,"prs":1409,"releases":4,"issues_open":246,"issues_closed":192,"prs_open":96,"prs_closed":355,"prs_merged":958,"discussions":3},"growth":{"day":{"stars":3,"forks":1,"base_day":"2026-11-15"}},"history":{"dates":["2026-11-10","2026-11-11","2026-11-12","2026-11-13","2026-11-14","2026-11-15","2026-11-16"],"stars":[5623,5632,5644,5644,5648,5651,5654],"forks":[836,838,842,842,842,844,845],"watchers":[null,17,17,17,17,17,17],"issues_open":[null,245,245,245,245,245,246],"prs_open":[null,103,99,98,95,95,96]},"boards":{},"same_owner":[],"issues":[{"number":537,"title":"[Troubleshoot]: CUDA Out of Memory error during training","comments":22,"created_at":"2026-01-08T02:14:01Z"},{"number":682,"title":"[Troubleshoot]: the rewards are always 0.0.","comments":21,"created_at":"2026-02-05T03:59:02Z"},{"number":267,"title":"[Troubleshoot]: RuntimeError: The tensor has a non-zero number of elements, but its data is not allocated yet.","comments":21,"created_at":"2025-11-13T06:51:39Z"},{"number":953,"title":"Hello, bash examples/embodiment/run_embodiment.sh libero_spatial_ppo_openpi_quickstart, the following errors occur: AttributeError: 'EGL . Thanks.","comments":14,"created_at":"2026-03-28T08:37:49Z"},{"number":655,"title":"[Troubleshoot]: Cannot resume from checkpoints","comments":13,"created_at":"2026-02-02T06:21:55Z"},{"number":470,"title":"[Troubleshoot]: The KL divergence of PPO is unstable","comments":12,"created_at":"2025-12-19T17:11:37Z"},{"number":259,"title":"[Bug]: 显卡进程混乱","comments":12,"created_at":"2025-11-12T04:41:03Z"},{"number":92,"title":"[Feature Request / Question] Support for NVIDIA RTX 5090 (Blackwell Architecture)","comments":12,"created_at":"2025-09-12T08:09:54Z"}],"discussions":[{"number":1764,"title":"Hello","comments":0,"category":"General","created_at":"2026-10-27T17:31:47Z"},{"number":481,"title":"Training time for RL in simulators","comments":3,"category":"Q&A","created_at":"2025-12-22T21:02:32Z"},{"number":329,"title":"Welcome to RLinf Discussions!","comments":0,"category":"Announcements","created_at":"2025-11-26T13:51:37Z"}],"latest_release":{"name":"RLinf v0.4 Release","tag":"v0.4","published_at":"2026-10-20T10:46:17Z","body":"🚀 RLinf v0.4 is here. Reinforcement learning steps out of simulation and into agile control of the physical world.\r\n\r\nv0.3 connected the path from simulation to real robots. v0.4 is the step toward deploying embodied RL at real-world scale. The release rebuilds the stack around three pillars: a modular real-robot system, high-efficiency real-robot algorithms, and full-modality models on a heterogeneous compute base.\r\n\r\n- 🧩 **Modular real-robot system** (`rlinf.robotics`): arms, dexterous hands, cameras, mobile bases, and teleoperation devices become composable blocks. Hardware drivers are fully decoupled from task logic, so changing a robot body, a configuration, or a cross-node deployment does not require changes to task or policy code.\r\n- ⚡ **Real-robot SOTA post-training**: representation-decoupled online RL (RLT), label-free self-supervised offline RL (STEAM), on-policy distillation (OPD) that reaches 96%+ success in 30 steps, and asynchronous real-time control (RTC) that compresses inference wait to about 2 ms. Together they address expensive real-robot samples, costly labels, and inference stutter.\r\n- 🌐 **Full-modality models**: a PyTorch-native OpenPI (π₀ / π₀.₅) rewritten to match official JAX numerics operator by operator; NVIDIA Cosmos3, FastWAM, π₀-FAST, Evo-1, and MolmoAct2; and the first extension of RL to SD3 / Wan2.2 visual generation.\r\n- 🤝 **Eight accelerator backends and the open ecosystem**: Kunlunxin, Biren, MetaX, and Intel XPU join the stack. Huawei Ascend and Moore Threads reach up to 2.78× throughput, with convergence aligned to A100. RLinf joins the PyTorch Ecosystem, ships as the RL training engine in NVIDIA Isaac Lab v3.0, and publishes joint practices with SGLang and AMD.\r\n\r\nTraining and evaluation recipes covered by this release are validated for correctness and reproducibility (see the test results at the end).\r\n\r\n### Embodied\r\n\r\n#### 1. Real World\r\n\r\nA unified robot interface decouples hardware from tasks. Arms, end-effectors, cameras, and bases compose freely. The release adds three real-robot platforms and a full dual-arm / single-arm collect–train–deploy toolchain.\r\n\r\n**Robot architecture:**\r\n\r\n- **Modular real-robot interface** (`rlinf.robotics`): drivers, task definitions, and environment wrappers are decoupled. Arms, dexterous hands / grippers, multi-view cameras, and mobile bases are addressed by one hierarchy. Swapping a driver or a cross-node topology does not require task or policy changes. Links: [Robot Interface](https://rlinf.readthedocs.io/en/latest/rst_source/concepts/robotics.html) ｜ [Robot Architecture](https://rlinf.readthedocs.io/en/latest/rst_source/concepts/robotics_architecture.html) ｜ [Add a Robot](https://rlinf.readthedocs.io/en/latest/rst_source/extending/new_robot.html)\r\n\r\n**Platforms and training pipelines:**\r\n\r\n- **Dual-arm Franka**: PICO VR bimanual teleoperation, per-arm HG-DAgger, and native PyTorch OpenPI π₀.₅ dual-arm SFT and deployment. Links: [Dual Franka PICO DAgger](https://rlinf.readthedocs.io/en/latest/rst_source/examples/embodied/dual_franka_pico_dagger.html) ｜ [Dual Franka OpenPI](https://rlinf.readthedocs.io/en/latest/rst_source/examples/embodied/dual_franka_openpi_pytorch.html)\r\n- **Single-arm Franka**: a Qwen-VL real-world reward model, depth capture, smooth intervention, online DAgger writes into LeRobot episodes, and LeRobot dataset merge / filter tools. Link: [Real-World Reward Model](https://rlinf.readthedocs.io/en/latest/rst_source/examples/embodied/franka_reward_model.html)\r\n- **New platforms**: dual-arm YAM (PICO VR teleoperation and LeRobot collection), AgileX Piper, and SO-101. Links: [AgileX Piper](https://rlinf.readthedocs.io/en/latest/rst_source/examples/embodied/piper.html) ｜ [SO-101](https://rlinf.readthedocs.io/en/latest/rst_source/examples/embodied/so101.html)\r\n\r\n#### 2. Algorithms\r\n\r\nNew SOTA algorithms cover online real-robot RL, label-free offline policy optimization, on-policy distillation, and high-frequency real-time control. They c","is_prerelease":false},"readme":true,"readme_zh":true,"news":[],"stale":false,"snapshot_at":"2026-11-16T09:23:14Z","history_start":"2026-09-06"}