Training humanoids to fail safely in digital twins before deployment

Niantic Spatial and Flexion report transferring an RGB-only local navigation policy from simulation to a real office without additional real-world training. The demonstration uses photorealistic Gaussian-splat rendering to support reinforcement learning rather than relying only on untextured geometric training scenes.

Digital twins could let robot developers run failure-heavy training and test site-specific behavior without exposing physical hardware to every mistake. For facility operators, the potential benefit is better preparation before robots enter a site, although this demonstration does not establish operational reliability.

Robots learning through reinforcement learning need repeated failures, but physical collisions and missteps can damage hardware and slow training. Niantic Spatial and Flexion present a simulation-first navigation demonstration designed to move that trial and error into a reconstructed environment. Their pipeline combines a digital twin of a real site, photorealistic Gaussian-splat rendering, NVIDIA Isaac Sim and Isaac Lab, and Flexion’s robot policies and deployment software. The companies report that an RGB-only local navigation policy trained entirely in simulation transferred zero-shot to a real office, with rendering fast enough to support training in hours. The change is not simply more simulation: it is using the visual appearance of an actual deployment environment rather than relying on randomized, untextured geometry. For site operators, this suggests a route to rehearsing robot behavior before deployment. However, the supplied account provides no quantitative success rates or evidence of broader deployment reliability.

Evidence supplied by Niantic Spatial, Inc..