FOCAL‑VLA is a framework that improves vision‑language‑action models by combining subtask‑guided geometry distillation with implicit world modeling. It transfers geometric knowledge from VGGT to focus on subtask‑relevant image regions and uses Track4World features to capture future 3D evolution, guiding action generation without running these models at inference time. Experiments demonstrate that FOCAL‑VLA outperforms baselines on both simulation benchmarks and real‑world manipulation tasks.
By Zhiyuan Gao, Di Wen, Yanxiang Zhan, Mohammad Khoshnazar, Jeroen Sch\"afer, Kunyu Peng, Michael Beetz
arXiv:2606. 24472v1 Announce Type: cross Abstract: Vision-language-action (VLA) models have made rapid progress in generalist robot manipulation by harnessing semantic knowledge from pretrained vision-language backbones, but their visual tokens remain grounded in 2D image coordinates rather than the calibrated geometry of the robot's cameras -- a mismatch especially pronounced in multi-camera setups, where views are coupled by known intrinsics and extrinsics yet processed as independent images.
By Yue Peng, Yongzhe Zhao, Artur Habuda, Khuyen Pham, Yanheng Zhu, Tran Nguyen Le, Fares Abu-Dakka, Li Guo
ForeAct3D is a framework that grounds future world modeling in Vision‑Language‑Action policies by decoding depth, semantic segmentation, and camera pose from policy representations into current and future 3D scene states. The future queries are conditioned on the policy‑generated action chunk, and a physical‑consistency closure enforces background staticity, instance‑level rigidity, and wrist‑camera pose alignment with end‑effector kinematics. This approach eliminates the need for future prediction at inference and achieves high success rates on LIBERO and CALVIN, outperforming the base policy across all suites.
By Zhe Tao, Feiran Wang, Gaowen Liu, Ramana Rao Kompella, Yan Yan
arXiv:2606. 17046v1 Announce Type: cross Abstract: Generalist robot policies must follow user instructions while reasoning about how objects, cameras, and robot actions interact in the 3D physical world.
By Jisang Han, Seonghu Jeon, Jaewoo Jung, Ren\'e Zurbr\"ugg, Honggyu An, Tifanny Portela, Marco Hutter, Marc Pollefeys, Seungryong Kim, Sunghwan Hong
arXiv:2608. 07267v1 Announce Type: new Abstract: Recent vision-language navigation (VLN) systems increasingly adapt pretrained vision-language models (VLMs) into vision-language-action (VLA) policies that map egocentric observations and language instructions directly to navigation actions.
By Yuehao Huang, Yunzi Wu, Xiaotao Zhang, Xinhai Li, Jiankun Dong, Jiajun Lv, Chi Zhang, Chenjia Bai, Yong Liu, Xuelong Li
arXiv:2606. 08530v1 Announce Type: cross Abstract: Vision-Language-Action (VLA) models achieve strong benchmark performance but still struggle in real-world deployment with unseen objects, background shifts, and different robot embodiments.
By Yuan Zhang, Shiqi Zhang, Yedong Shen, Shuai Dong, Jiajun Deng, Xin Zhang, Yuxuan Gao, Jiajia Wu, Xin Nie, Zhiyuan Cheng, Jianmin Ji, Yanyong Zhang, Xingyi Zhang, Jia Pan
arXiv:2609.25558v1 Announce Type: cross
Abstract: Vision-language-action policies benefit from geometric supervision, but current-frame geometry alone does not explicitly describe the changes associa...
By Jinu Pahk, Jesoon Kang, Taegeon Park, Jisu An, Soo Min Kimm, Jaejoon Kim, Byoung-Tak Zhang
arXiv:2605. 21862v2 Announce Type: replace-cross Abstract: Chunked vision-language-action (VLA) policies predict multi-step robot controls, conditioning each update on the current visual observation alone.
By Chushan Zhang, Ruihan Lu, Jinguang Tong, Xuesong Li, Yikai Wang, Hongdong Li
GeoScaffold introduces a geometric supervision framework that embeds 3D geometry into vision‑language navigation policies during training, eliminating the need for depth sensors or 3D encoders at inference. The method learns a compact depth tokenizer from training trajectories, then fine‑tunes the policy with geometry query tokens that reconstruct depth, connectivity, and traversability, turning these states into compact geometric latents for action decoding. After training, the tokenizer, generators, and reconstruction heads are discarded, leaving a lightweight backbone that outperforms vision‑only navigators on continuous VLN benchmarks.
By Yixuan Jiang, Wentong Li, An Liu, Zihao Xin, Fulin Tang, Cong Leng, Yang Gao, Jian Cheng
Mem-World introduces a memory‑augmented action‑conditioned world model for robot manipulation, featuring W‑VMem—a 4D wrist‑view‑centered surfel‑indexed memory that anchors historical observations to evolving surface elements. By explicitly modeling when and where scene elements are observed, the system retrieves geometry‑aware history frames during generation, providing informative, non‑redundant context for future action predictions. Experiments demonstrate that Mem‑World produces persistent rollouts, improves policy evaluation reliability (14.5 % higher Pearson correlation with real‑world performance), and boosts long‑horizon task success rates from 58 % to 72 % using synthetic data generation.
By Zirui Zheng, Jiaqian Yu, Xiongfeng Peng, jun shi, Mingyi Li, Chao Zhang, Weiming Li, Dong Wang, Huchuan Lu, Xu Jia
The paper introduces IG‑VLA, a vision‑language‑action framework that learns to imagine task‑relevant future scene evolution in a latent spatiotemporal space, guiding action prediction without generating full pixel‑level videos. It further compresses this future reasoning into a compact Scene Gist Token via Scene Gist Memory, enabling efficient inference. Experiments on LIBERO, LIBERO‑Plus, and VLABench show that IG‑VLA improves success rates by nearly 6% and speeds up inference up to 6.38× compared to strong baselines.
By Shenglan Li, Zhendong Mi, Hengyi Zhu, Jingwu Luo, Chun Kit Chan, Geng Yuan, Yanzhi Wang, Pu Zhao, Shaoyi Huang
GeoLAM is a framework that learns geometry‑grounded latent actions from unlabeled human videos. It uses future‑frame reconstruction with a frozen geometric feature hierarchy and motion supervision from a 4D geometry teacher to capture 3D displacement, image‑plane motion, and surface‑orientation changes. After pretraining, the representation serves as transition targets for a world‑action model trained on robot demonstrations, enabling denoised latent actions and executable action chunks without requiring hand‑pose annotations or future‑video generation during deployment.
By Yifan Xie, Hekun Tian, Jinkun Liu, YuAn Wang, Qiao Sun, Wenbo Ding