RLinf v0.3 wraps robot training into one open pipeline

RLinf started as a research reinforcement learning system for robots. Its third release, v0.3, according to the project's GitHub page, turns it into something closer to a full pipeline.

Infinigence, a Chinese AI infrastructure company, and Tsinghua University shipped v0.3 on July 16, framing it as a one-stop development platform for embodied AI, the field of training AI systems, often robots, to act in physical environments rather than only process text. The release wraps data collection, supervised fine-tuning (SFT), reinforcement learning, evaluation, and real-robot deployment under one stack. It is the first version in the project's trajectory to do so.

v0.1 covered only the reinforcement learning system abstraction, the underlying training machinery without the rest of the pipeline. v0.2 added real-world online learning infrastructure. v0.3, the project's own release notes say, is where RLinf becomes "a one-stop development platform for embodied AI that closes the simulation-to-real-and-back loop."

That loop is the practical problem. Embodied AI training has historically required stitching together separate tools: one for collecting demonstration data, another for fine-tuning a model on that data, a third for reinforcement learning, and a fourth for moving the trained policy onto a robot. Each handoff is a place where data formats, simulators, and hardware drivers can fail to align. v0.3's pitch is to remove those handoffs.

v0.3 adds six embodied models: Dexbotic DM0 with online RL fine-tuning on the LIBERO benchmark; DreamZero, a vision-language-action (VLA) policy fine-tuned inside a video-generation world model; GR00T-N1.6 and N1.7 for RL fine-tuning; ABot-M0; StarVLA, trained with GRPO (Group Relative Policy Optimization) on LIBERO; and LingBot-VLA for SFT and RL on RoboTwin. The project says DreamZero's training throughput improves roughly 4x. That number comes from the project's own reproducibility tests documented in the v0.3 Release Notes, not from independent benchmarks.

v0.3 extends DSRL to the Pi0.5 VLA model, adds the RECAP offline advantage-estimation pipeline, and brings SAC-Flow to the DOS-W1 robot. On the simulation side, the release adds an asynchronous PPO (Proximal Policy Optimization) configuration and offline IQL (Implicit Q-Learning) on the D4RL benchmark suite. For human-in-the-loop learning, v0.3 ships DAgger (Dataset Aggregation), an online imitation learning method, and HG-DAgger for real-machine training.

Real-robot support is where v0.3 tries hardest to close the loop. The release adds three teleoperation methods (Spacemouse, VR, and GELLO), three new robot platforms (a dual-arm Franka with both joint-space and task-space control, the GimArm, and the DOS-W1), and two end-effectors (the Franka DexHand dexterous hand and a Franka Robotiq gripper). The release also adds LeRobot-format data support, SFT deployment for Pi0 on real hardware, and a pipeline for collecting and annotating reward-model data.

The project is publishing v0.3 in coordination with multiple cloud providers, including Baidu Smart Cloud, which gives the release a deployment path beyond a researcher's own GPU cluster. The full Tech Report is available on arXiv, and the project README is public.

The qualifications matter. The "world's first" framing in the project's coverage is the project's own positioning, not an independent ranking. The SOTA real-machine task success-rate and ~4x DreamZero throughput numbers are project-reported reproducibility results, not external benchmarks. No customer adoption data, no third-party evaluation, and no independent analyst assessment is in the receipts. RLinf v0.3 is a real release of real open-source code, and the consolidation story is concrete: v0.1 covered RL systems, v0.2 added online learning, and v0.3 wraps the full data-to-deployment loop. What the project has not yet shown is that the unified stack produces better robot policies than assembling the same components from existing tools. That is the next test.