Niantic Spatial and Flexion report sim-to-real navigation using an office digital twin

Niantic Spatial and Flexion report an end-to-end demonstration of RGB-only local navigation trained in a reconstructed environment and transferred zero-shot to a real office. The workflow combines Gaussian-splatting rendering with NVIDIA Isaac Sim and Isaac Lab.

Training inside a deployment-site digital twin could let facility operators test navigation without exposing physical robots to every training failure. The reported office demonstration does not establish reliability across other sites or more complex tasks.

Niantic Spatial and Flexion describe a pipeline that connects camera-based site capture to robot training in NVIDIA Isaac Sim and Isaac Lab. Niantic Spatial reconstructs the environment, while Flexion supplies policies and deployment software tuned to the robot hardware. The companies report that an RGB-only local navigation policy trained entirely in simulation, using Gaussian-splatting rendering, transferred zero-shot to a real office. They also report that rendering speed and quality supported reinforcement-learning training in hours. The shift is toward training with realistic imagery of a deployment site rather than relying on randomized, untextured geometry and depth inputs. For operators, this points to a way to rehearse navigation and adapt robots to specific facilities before physical deployment. However, the supplied excerpt provides no quantitative success rates or cross-site validation; broader reasoning and manipulation capabilities remain future ambitions.

Evidence supplied by Niantic Spatial, Inc..