🤖We are hiring multiple Summer'26 Research Interns at @amazon FAR to work on open-world navigation and robot foundation models, especially in neural rendering & simulation/predictive world models/reasoning & agency/real-world evaluation/long-term autonomy!
What if robots could interact with the real world before ever stepping into it? 🤖
Wanderland (Liu et al., CVPR 2026 Highlight) from NYU AI4CE addresses real-world exploration. It captures real New York streets, lounges and rooftops as dense, metric 3D: point clouds, Gaussian splats, meshes and LiDAR.
In collaboration with the Wanderland team @simbaforrest@xinhao6iu , we used Nova to convert 20 of these captures into interactive, simulation-ready environments.
Nova generates each environment automatically as an executable scene program, from the capture images alone (10 scenes) or from the images and the LiDAR mesh (10 scenes). No manual modeling or annotation is involved, and every annotation is exact by construction.
What this provides for robot learning:
• Real-to-sim environments: training layouts taken from real urban sites, not synthetic rooms
• Interactive dynamics: 8,788 objects, including 242 articulated doors, gates and lids (revolute and prismatic joints with axes and limits) and 351 free-moving bodies with mass, friction and convex collision geometry
• Controlled variation: each scene is a parameterised program, which enables domain randomization and single-factor ablations within the same real site
• Ground-truth annotations: instance and semantic labels, metric 3D bounding boxes, joint parameters, and physical properties with their provenance, for supervision and task-success evaluation
• Cross-simulator evaluation: native exports for Isaac Sim, MuJoCo, Gazebo and PyBullet
Nova datasets: nova3d.xyz/datasets
Wanderland: ai4ce.github.io/wanderland/#EmbodiedAI#Robotics#RobotLearning#Simulation#Real2Sim#CVPR2026
Fun clip in my #IROS2026 talk from Mission: Impossible – Fallout as a checklist for Embodied Spatial Intelligence: navigate, follow directions, avoid people, manipulate objects, jump out a window 😂 To achieve this, Robots must couple world + self through action, time, & memory.
We are hiring Summer 2026 Research Interns at @amazon Frontier AI & Robotics (FAR) to work on open-world navigation in the era of robot foundation models and Internet-scale data! We are especially interested in candidates with strengths in robotic navigation, multi-modal foundation models, reasoning and agency, neural rendering, or real-world evaluation.
If interested, please email me (ystian@amazon.com) and @simbaforrest (nycfeng@amazon.com) with the following subject line: [FAR Intern - Navigation] Your full name + School, attach your CV and your maximum availability window, along with a brief note about your background and interests.
If excited, please email both @YunshengTian (ystian@amazon.com) and me (nycfeng@amazon.com) with the subject line: [FAR Intern - Navigation] Your full name + School.
Do attach your CV and your maximum availability window, and a brief note about your background and interests.
A robot that can navigate and vacuum your living room may still be lost in the real world. True open-world navigation is not just a problem of geometry: it requires spatial understanding, memory, and reasoning at scale.
Have you ever been lost in a new place and wished a robot could reliably guide you? Or waited for a sidewalk delivery bot, only to learn that it got confused, stuck, or gave up halfway?
EgoPush
A learning framework that enables mobile robots to perform long-horizon multi-object rearrangement using only egocentric vision—no global maps or external tracking needed. Uses object-centric latent states and privileged RL teacher distilled into a visual student, with zero-shot sim-to-real transfer.
No matter whether you are with NYU or not, please feel free to contact us and join this fun open-source effort! It will pull you out of your coach and walk you in your city🚶♂️➡️, so let's scan more places to advance robotics and embodied AI together!
Our amazing students and researchers at NYU Center for Robotics and Embodied Intelligence (CREO) and our collaborators scanned 500+ outdoor & indoor places with high fidelity, large scene diversity, and scale, covering totally about 4 million square meters, and 400K+ frames.
To advance embodied AI, we need more high-quality photo-realistic and geometry-realistic simulation environments, especially for reproducible closed-loop evaluation. Wanderland is our first step to address this urgent need. Join us in this open-source effort and scan more places!
🌍 Is YouTube + VGGT + 3DGS enough for simulative environments?
🧩 What’s missing in Real2Sim in 2026?
🔁 How do we evaluate embodied agents in a reliable closed-loop?
🔮 Meet Wanderland: a Real2Sim framework, benchmark, dataset, and environment for mobile agents!
[1/n]