Qwen-RobotNav is a scalable navigation model with a configurable observation protocol, unifying instruction following, point- and object-goal navigation, target tracking, autonomous driving, and embodied question answering in an agentic navigation system.

Qwen-RobotManip is a generalizable VLA foundation model that aligns heterogeneous manipulation data across representation, motion, and behavior, enabling large-scale cross-embodiment training and strong out-of-distribution generalization.

Qwen-RobotWorld is a language-conditioned video world model that predicts physically grounded futures across manipulation, autonomous driving, indoor navigation, and human-to-robot transfer through a unified natural-language action interface.

Qwen-VLA is a unified embodied foundation model that casts manipulation, navigation, and trajectory prediction into a shared action-and-trajectory prediction framework, with one model operating across tasks and robot embodiments.

We present MM-Nav, a multi-view Vision-Language-Action (VLA) system with 360° observation, built upon pretrained large language models and visual foundation models. The model is trained on large-scale expert navigation data collected from three reinforcement learning (RL) agents.

Habitat-GS extends Habitat-Sim with real-time 3D Gaussian Splatting scene rendering and drivable Gaussian avatars, providing photorealistic, dynamic, and Habitat-compatible environments for embodied navigation.

LDA-1B is a dynamics-centric robot foundation model trained on more than 30,000 hours of heterogeneous embodied data, jointly learning latent dynamics, visual forecasting, and policy in a unified framework.

UrbanVLA is a route-conditioned Vision-Language-Action model for urban micromobility. It aligns noisy navigation-tool routes with visual observations to enable scalable, long-horizon navigation. Trained via a two stage pipeline including SFT and RFT, UrbanVLA outperforms baselines by over 55% on MetaUrban and achieves robust real-world navigation across 500m+ routes.

TrackVLA++ is a novel Vision-Language-Action model that incorporates spatial reasoning and target identification memory, enabling superior performance in both long-horizon and highly crowded tracking scenarios.

We introduce a cross-embodiment and cross-task Navigation Foundation Model (NavFoM), trained on eight million navigation samples that encompass quadrupeds, drones, wheeled robots, and vehicles, and spanning diverse tasks such as vision-and-language navigation, object searching, target tracking, and autonomous driving.

OctoNav introduces a generalist embodied navigation benchmark and VLA model for following free-form instructions that can combine arbitrary modalities and navigation capabilities.

DreamVLA forecasts compact dynamic, spatial, and semantic world knowledge to form a perception-prediction-action loop for robotic manipulation, improving both reasoning and action generation.

TrackVLA is a vision-language-action model capable of simultaneous object recognition and visual tracking, trained on a dataset of 1.7 million samples. It demonstrates robust tracking, long-horizon tracking, and cross-domain generalization across diverse challenging environments.

LaDi-WM predicts future geometric and semantic latent states with diffusion modeling and uses the imagined states to guide a predictive manipulation policy in simulation and the real world.

OnlineAnySeg uses voxel hashing and visual-foundation-model features to merge sequential 2D masks into online zero-shot 3D instances in real time.

CogNav models ObjectNav as an LLM-directed cognitive process over an online heterogeneous map, enabling efficient, human-like object search in unseen environments.

We present Uni-NaVid, the first video-based vision-language-action (VLA) model designed to unify diverse embodied navigation tasks and enable seamless navigation for mixed long-horizon tasks in unseen real-world environments.

NaVid makes the first endeavour to showcase the capability of VLMs to achieve state-of-the-art level navigation performance without any maps, odometer and depth inputs. Following human instruction, NaVid only requires an on-the-fly video stream from a monocular RGB camera equipped on the robot to output the next-step action.

We present Open6DOR, a challenging and comprehensive benchmark for open-instruction 6-DoF object rearrangement tasks. Following this, we propose a zero-shot method, Open6DORGPT, which achieves SOTA performance and proves effective in demanding simulation environments and real-world scenarios.

We propose a robust zero-shot 3D instance segmentation method that leverages the 3D view consensus of 2D candidate masks. Our method can integrate with a 2D visual foundation model (e.g., CLIP) to achieve open-vocabulary 3D instance segmentation.

We propose a graspability-aware mobile manipulation approach powered by an online grasping pose fusion framework that enables a temporally consistent grasping observation.

We introduce MIPS-Fusion, a robust and scalable online RGB-D reconstruction method based on a novel neural implicit representation – multi-implicit-submap.

We propose a framework for the challenging 3D-aware ObjectNav based on two straightforward sub-policies, namely corner-guided exploration policy and category-aware identification policy.

To realize efficient random optimization in the 18D state space of IMU tracking, we propose to identify and sample particles from active subspace.

We propose to tackle the difficulties of fast-motion camera tracking in the absence of inertial measurements using random optimization.

We propose a novel fusionaware 3D point convolution which operates directly on the geometric surface being reconstructed and exploits effectively the inter-frame correlation for high quality 3D feature learning.

We propose a novel approach to robot-operated active understanding of unknown indoor scenes, based on online RGBD reconstruction with semantic segmentation.