Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
UP-VLA: A Unified Understanding and Prediction Model for Embodied Agent
Authors: Jianke Zhang, Yanjiang Guo, Yucheng Hu, Xiaoyu Chen, Xiang Zhu, Jianyu Chen
ICML 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results show that UP-VLA achieves a 33% improvement on the Calvin ABC-D benchmark compared to the previous state-of-the-art method. Additionally, UP-VLA demonstrates improved success rates in real-world manipulation tasks, particularly those requiring precise spatial information. |
| Researcher Affiliation | Academia | 1Institute for Interdisciplinary Information Sciences, Tsinghua University,Beijing, China. 2Shanghai Qi Zhi Institute, Shanghai, China. Correspondence to: Jianyu Chen <EMAIL>. |
| Pseudocode | No | The paper describes the methodology using textual explanations and diagrams (e.g., Figure 4) but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code can be found at https://github.com/ Claderny Jorn/UP-VLA. |
| Open Datasets | Yes | For simulation evaluation, we utilize CALVIN (Mees et al., 2022), an open-source benchmark to learn long-horizon language-conditioned tasks. We mix training data across two domains: one part is from Bridge (Walke et al., 2023), which includes 25k robotic arm demonstrations. Another part is from LLava-tuning-665k (Liu et al., 2024), which includes 665k image-text pairs. |
| Dataset Splits | Yes | For the simulation environment data, we follow the setups of the CALVIN benchmark (using its training sets and evaluation sets). |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., CPU, GPU models, memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions using models like Show-o (Xie et al., 2024), CLIP-ViT (Radford et al., 2021), MagVIT (Yu et al., 2023), and VQ-GAN (Esser et al., 2021) but does not specify software dependencies with version numbers (e.g., Python, PyTorch versions). |
| Experiment Setup | Yes | In the pretrain stage, we train UP-VLA for 20k steps with batch size of 64 on future prediction and vision-language understanding tasks. We apply a linear warmup at the first 1k steps. In the action learning stage, we train UP-VLA with a batch size of 64. We initialize the backbone of UP-VLA using Show-o (Xie et al., 2024). During training, we fully fine-tune the parameters of the LLM and freeze all encoders. We use varying weights to combine these three losses: L = λ1LMMU + λ2LP RE + λ3LACT |