Example Gallery#
This section presents the collection of examples currently supported by RLinf, showcasing how the framework can be applied across different scenarios and demonstrating its efficiency in practice. This example gallery is continuously expanding, covering new scenarios and tasks to highlight RLinf’s flexibility and efficiency.
Embodied intelligence is RLinf’s primary focus. The embodied gallery is split into seven entry points — pick the one that matches your starting question:
Start from a benchmark — LIBERO, ManiSkill, RoboTwin, IsaacLab, and more.
Run on physical robot hardware — the Franka family plus GimArm, XSquare Turtle2, and DOS-W1.
RL-fine-tune embodied model families — π₀, GR00T, StarVLA, Lingbot-VLA, and more.
Train and fine-tune world models — OpenSora, Wan, and more.
Build reward models for simulation and real-world reinforcement learning workflows.
Supervised fine-tuning recipes that produce strong RL cold-start checkpoints.
Algorithm-centric examples — DAgger, RECAP, DSRL, IQL offline RL, sim-real co-training, MLP / SAC-Flow.
Beyond embodiment:
Fine-tune SD3 and Wan2.2 image/video generation models with Diffusion-NFT.
Math reasoning and agentic AI workflows, in both single-agent and multi-agent settings.
Flexible, dynamic scheduling of compute resources across the most suitable hardware devices.