Event

RLinf v0.4 released, upgrading real-robot training and domestic compute support

1 reports · 1 sources

RLinf, an open-source embodied reinforcement learning training framework developed by Tsinghua University, Infinigence AI, together with Beijing Zhongguancun Academy and other institutions, has officially released v0.4. It iterates around four directions: a LEGO-style real-robot system, real-robot post-training algorithms, full-modality model support, and a domestic compute foundation, further focusing on reinforcement learning needs in real physical scenarios. v0.4 upgrades by default to PyTorch 2.11 and new versions of SGLang / vLLM. The official guidance requires using requirements/install.sh to rebuild the virtual environment or pulling the rlinf0.4 series Docker images, and not directly reusing the v0.3 virtual environment; if Checkpoints trained with the older OpenPI-PyTorch need to be loaded, they must first be converted with the Checkpoint conversion tool.