機器人與具身 AI
VLANeXt Expands Into a Unified VLA Research Foundation, Covering World Models and Latent Action Pretraining
A major VLANeXt library update brings World Action Models, JEPA-style world modeling, and latent action learning into a single training and evaluation framework. The team reports an average success rate of 83.9% on LIBERO-plus, although the results still require independent reproduction across libraries and on physical robots.
On August 31, the VLANeXt team announced a major library upgrade, expanding the research code originally built to disentangle vision-language-action model design into a unified foundation for comparing multiple robot learning approaches. The new version supports World Action Models, latent action pretraining and fine-tuning, JEPA-style world modeling, small-scale VLAs, language-action learning, additional data formats, and parallel evaluation. The original paper version remains available on the `VLANeXt-ori` branch, preventing the new abstraction layers from disrupting the reproduction path for existing results.
VLANeXt was designed to separate choices that are easily conflated in VLA systems while keeping training and evaluation conditions consistent. The team’s twelve documented recipes include dividing responsibilities between the VLM and a dedicated policy module, predicting an action chunk at once, representing actions as continuous values rather than discrete tokens, using a stronger vision-language backbone, adopting a soft connection between the VLM and policy, and conditioning on robot proprioception. Historical frames do not necessarily deliver consistent improvements; world models may help, but they substantially increase computational cost. These options can now be switched within shared configurations and data pipelines, making the framework better suited to ablation studies than stitching together separate research codebases one by one.
Under LIBERO-plus perturbations involving cameras, lighting, backgrounds, language paraphrases, robot states, and scene configurations, the project reports an average success rate of 83.9% for VLANeXt, compared with 69.6% for OpenVLA-OFT. Training can use the reorganized and filtered MolmoAct2-DROID dataset, while evaluation covers both LIBERO and LIBERO-plus.
The significance of this update lies in its research infrastructure, not merely its leaderboard position: a shared foundation can reduce comparison bias caused by differences in data processing, policy heads, and evaluation scripts. However, the figures were produced by the authors under their own setup, and the data reorganization may also affect the fairness of comparisons. LIBERO success rates do not directly indicate the safety, long-term reliability, or cross-hardware generalization of physical robotic arms. Adopters should also note that the project uses the NTU S-Lab License; commercial use should not be assumed to be equivalent to Apache 2.0.