NVIDIA has introduced Isaac GR00T N1.6, a next-generation multimodal vision-language-action model designed to enable generalist capabilities in humanoid robots through a comprehensive sim-to-real workflow. The system integrates advanced perception, planning, and control to support complex tasks in dynamic, real-world environments. At the core of GR00T N1.6 is a vision-language-action (VLA) architecture that unifies egocentric visual input, robot state data, and natural language instructions into a single policy. It leverages world models like NVIDIA Cosmos Reason to break down high-level tasks into step-by-step action plans grounded in real-time scene understanding. This allows the robot to perform both locomotion and dexterous manipulation through end-to-end learned representations. Key improvements in GR00T N1.6 include enhanced reasoning and perception through a refined version of the Cosmos-Reason-2B vision-language model with native resolution support, enabling clearer visual input and more accurate environmental understanding. The model also features a 2x larger diffusion transformer with 32 layers and state-relative action prediction, resulting in smoother, more fluid, and adaptive motion. Additionally, training on a broader range of teleoperation data from diverse robot types—including humanoids, mobile manipulators, and bimanual arms—improves cross-embodiment generalization. Pretrained weights are available for zero-shot evaluation of basic manipulation skills, with further fine-tuning recommended for specific hardware or tasks. The system’s low-level control is powered by whole-body reinforcement learning (RL) trained in NVIDIA Isaac Lab and Isaac Sim. These policies generate human-like, dynamically stable motion for locomotion, manipulation, and multi-contact interactions. The sim-to-real transfer is achieved with zero-shot deployment, meaning the policies work directly on physical robots without additional tuning, significantly reducing development time and effort. For navigation, GR00T N1.6 is enhanced using synthetic data generated by COMPASS, a novel workflow that combines imitation learning, residual RL, and policy distillation. COMPASS creates diverse, high-quality navigation trajectories across varied environments and robot types, enabling the model to learn point-to-point navigation in simulation. The navigation policy outputs velocity commands to the whole-body controller, which handles balance and contact dynamics, while the high-level policy focuses on path planning, obstacle avoidance, and task handoffs. This approach enables zero-shot sim-to-real transfer, even in new physical spaces, without collecting real-world data. To ensure accurate positioning in large-scale environments, GR00T N1.6 employs a vision-based localization system built on NVIDIA CUDA-accelerated libraries. The system uses stereo cameras to create and maintain prebuilt maps, including a cuVSLAM landmark map, a cuVGL bag-of-words map, and an occupancy map with semantic labels. At runtime, cuVGL identifies similar image pairs from the map to estimate an initial pose, which cuVSLAM refines by matching local features and performing continuous tracking and map optimization. This results in low-drift, real-time localization. The full stack is supported by NVIDIA Isaac ROS, with tools for offline map creation and real-time localization pipelines using stereo cameras, image rectification, and CUDA-accelerated SLAM modules. Developers can get started with NVIDIA Isaac libraries, access detailed documentation, and explore free Robotics Fundamentals courses. The GR00T N1.6 PointNav example provides code and instructions for replicating and extending the navigation stack. Stay updated via NVIDIA Robotics channels and join the developer community to advance physical AI systems.