AgiBot open-sources VINE, a value-gradient fine-tuning method for flow-matching policies
AgiBot released the VINE codebase on GitHub under Apache License 2.0. VINE enables stable end-to-end value-gradient fine-tuning of flow-matching policies by reconstructing interpolation states at each denoising step. The repository targets Linux, Python 3.12, CUDA 12, JAX 0.5.3, and OGBench 1.2.1.
China context
- Original name
- 智元机器人
- Outside China
- Open weights · github.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can access the VINE codebase on GitHub and use it under Apache 2.0, enabling experimentation with value-gradient fine-tuning for flow-matching policies without licensing restrictions.
- For investors
- The release signals AgiBot's continued investment in open-source robotics research, which may influence the competitive landscape for embodied AI tooling.
AgiBot published the VINE (Value-gradient Iterative Noise Exploration) repository, which implements a method for stable end-to-end value-gradient fine-tuning of flow-matching policies. Instead of following a single denoising trajectory, VINE reconstructs a fresh interpolation state at every denoising step, creating a stable differentiable path for value-gradient backpropagation while preserving the original flow-matching semantics. The code is built on top of FQL and is licensed under Apache License 2.0.
VINE addresses instability in value-gradient fine-tuning of flow-matching policies by reconstructing interpolation states at each denoising step, enabling stable backpropagation. The validated environment is Linux, Python 3.12, CUDA 12, JAX 0.5.3, and OGBench 1.2.1, with GPU required for experiments. The repository includes scripts for dataset download and experiment reproduction, with preset defaults for single-file datasets and overrides for 100M benchmarks.
Robotics and embodied AI developers outside China gain access to a new open-source method for fine-tuning flow-matching policies with value gradients, potentially reducing the engineering effort needed to stabilize such training. The Apache 2.0 license permits commercial use, which may lower barriers for companies building robot control systems.
The open-source release under Apache 2.0 allows companies to incorporate VINE into proprietary robot learning pipelines without licensing fees. The method's focus on stable fine-tuning could reduce development time for teams working with flow-matching policies.
A specific signal to watch is whether independent researchers reproduce VINE's results on OGBench benchmarks and report comparable performance. Another observable is adoption of the method in downstream robotics projects or integration into existing RL frameworks.