Rethinking the Practicality of Vision-language-action Model: A Comprehensive Benchmark and An Improved Baseline

A research consortium has published a paper, "Rethinking the Practicality of Vision-language-action Model," that critically assesses current VLAs and introduces a new solution. They argue existing models are hindered by excessive parameters, prohibitive pre-training costs, and limited embodiment diversity. To address this, they first created CEBench, a comprehensive new benchmark spanning diverse robotic embodiments in both simulation and the real world, supported by 14.4k simulated and 1.6k real-world expert trajectories. Using this testbed, they identified key practicality flaws and built LLaVA-VLA as an improved, deployable baseline.

LLaVA-VLA's architecture integrates a compact Vision-Language Model (VLM) backbone with specialized components for robotics: multi-view perception, proprioceptive tokenization (encoding the robot's own state), and action chunking. Crucially, it adopts a two-stage training paradigm of post-training and fine-tuning on CEBench, eliminating reliance on massive, costly pre-training datasets. The model also extends the action space to unify navigation and manipulation tasks. Experiments confirm its generalization across embodiments, and real-world tests establish it as the first end-to-end VLA capable of mobile manipulation, all while being lightweight enough to run on consumer-grade GPUs. The team will open-source all datasets, code, and model checkpoints to foster reproducibility.