Beyond Data Scaling: VLAct Makes Representation the VLA Priority

General-purpose vision-language-action models are often discussed through the lens of data scaling. Yet robot data does not behave like web-scale image-text data: collecting trajectories requires physical hardware, controlled environments, and costly human or autonomous operation. Coverage of the physical world is also sparse. Under a fixed trajectory budget, the central question is therefore not only how much data a model sees, but how much transferable knowledge it extracts from that data.

VLAct, featured by Hugging Face Daily Papers, takes a representation-centric approach to continued pre-training. The method first trains a VLA-oriented vision-language model backbone on broad and heterogeneous data collected across robot embodiments, and only then adapts it to downstream tasks. The goal is to avoid turning a general VLM into a narrow action-fitting system. Instead, VLAct attempts to retain the model’s visual-language capabilities while building action concepts that can be reused across different robots.

Its main design choices are:

• VLM-prior preservation: continued pre-training is designed to retain the broad visual and linguistic knowledge of the original backbone rather than allowing limited robot data to overwrite it.
• Multi-head continuous-action co-supervision: several action heads jointly learn continuous control signals, helping the model accommodate variation across datasets and embodiments.
• A partially unified action layout: action representations are aligned across embodiments where useful, while task-specific action heads remain available during fine-tuning.

This is a practical compromise between full action-space unification and completely separate policies. Different robots do not need identical control interfaces to share useful knowledge. They can still benefit from common structures such as grasping, moving, and object manipulation, before translating those structures into embodiment-specific joints or control dimensions.

The reported evaluation spans simulation, real-world settings, and transfer to an unseen embodiment. VLAct reaches success rates of 82.6% on LIBERO-Plus and 92.5% on RoboTwin 2.0, surpassing the industrial VLA systems listed in the paper summary, ABot-M0 and LingBot-VLA. On RoboDojo, it ranks sixth among all policies by success rate and outperforms the entries explicitly designated as world-action models. The strongest data-efficiency signal appears on RoboCasa-GR1, an unseen humanoid embodiment: using only 20% of downstream trajectories, VLAct outperforms the full-data GR00T-N1.6 baseline.

These results do not make data scale irrelevant. Rather, they suggest that representation quality is an independent axis of progress when data is scarce or expensive. The project also releases its data, models, and complete training and fine-tuning pipeline. According to the supplied material, full continued pre-training requires only 16 GPUs, which could make cross-embodiment experiments more accessible to research teams without industrial-scale infrastructure.

The broader implication is a shift in how VLA systems may be compared. Dataset size remains important, but the organization of action knowledge, retention of general VLM capabilities, and alignment across embodiments may determine how efficiently a model converts trajectories into downstream behavior. At the same time, the available summary does not establish the isolated contribution of every component, the exact data mixture, or robustness across a wider range of deployment conditions. VLAct is best viewed as strong evidence for representation-centric VLA training, not as a universal guarantee for every robot task.

Source: Hugging Face Daily Papers