Milestones

2026.10

v0.4 focuses on RL in the physical world. It iterates on four fronts, LEGO-style real-robot systems, efficient real-robot algorithms, support for omni-modal models, and a compute foundation built on Chinese chips, reshaping how real-robot RL is engineered and trained, and moving embodied RL from proofs of concept toward real robot applications that are more efficient, more flexible and reproducible.

2026.07

v0.3 upgrades RLinf across models, algorithms, real robots, simulation and systems, turning it into a one-stop platform for continually evolving embodied AI. For the first time it connects data collection, data management, supervised fine-tuning (SFT), reinforcement learning (RL), model evaluation and real-robot deployment, closing the loop from training in simulation to online learning on real robots and continued iteration.

2026.05

RLinf passed the PyTorch community's full review and meets the technical and quality standards of the PyTorch Ecosystem, an official recognition from the global AI developer community.

2026.05

The latest Agentic AI Landscape from Ant Open Source, a map of China's open-source ecosystem, groups projects into three layers, models, model infrastructure and agent infrastructure. RLinf, core RL infrastructure for the agent era, is now listed in the Model Infra layer.

2026.05

RLinf now supports the world action model DreamZero, and goes beyond adapting it. Building on its system-level optimizations, RLinf rebuilt and accelerated the DreamZero training pipeline, reaching nearly 4× the training throughput of DreamZero's official baseline scripts with better convergence. This extends RLinf from RL training of large models to training foundation models for embodied AI, and widens its ecosystem.

2026.05

RLinf-VLA is a unified, efficient framework for RL on vision-language-action (VLA) models, built to cover more, run faster and train better. More: one interface supports many VLA models, RL algorithms and heterogeneous simulators, which makes it easier to extend and to evaluate consistently. Faster: flexible resource allocation for the RL pipeline and a hybrid fine-grained pipelining strategy for GPU-parallel simulation speed up training by 1.61× to 1.88×. Better: models trained with it improve by about 20% to 85% on LIBERO, ManiSkill and RoboTwin, and the work distills a set of effective training practices. RLinf-VLA gives VLA RL research in embodied AI a unified, efficient and reproducible foundation.

2026.05

BEHAVIOR, the benchmark initiated by Prof. Fei-Fei Li, gives long-horizon embodied tasks a high-fidelity training ground, but also makes training slow and expensive. RLinf optimized BEHAVIOR at the system level, cutting end-to-end latency from 1028 ms/step to 41 ms/step, a 25× speedup.

2026.03

ModelScope, together with the CCF Technical Committee on Intelligent Robotics, the MIIT Key Laboratory of Equipment Digital Twin Technology, the Lumina embodied AI community and other institutions, published EAI-100 (Embodied Artificial Intelligence 100), the annual list of 100 representative results and people in embodied AI. RLinf was named one of its ten breakthrough projects.

2026.03

v0.2 takes RLinf into the real world with fully heterogeneous, fully asynchronous real-robot RL. An adaptive communication plane and a unified hardware interface let you use robots the way you use GPUs, and run real-world RL across domains and machines. Asynchronous training and inference support at least four data-efficient RL algorithms and human-in-the-loop training, giving scalable real-world RL a system-level foundation.

2025.09

v0.1 brings simulation, training and inference into one pipeline, with unified support for a wide range of embodied foundation models, RL algorithms and simulators. System-level optimizations put RLinf well ahead of existing frameworks in GPU utilization and overall training throughput, with up to 1.27× higher end-to-end efficiency in our experiments.