RobotPan Gives Humanoid Robots Full 360° Vision With 11x Fewer Primitives

Teleoperating a humanoid robot through a headset is often a nauseating experience. The camera view is narrow, the head bobs with every step, and switching between the half-dozen cameras bolted to the chassis breaks the operator's flow. A new system called RobotPan, built by the X-Humanoid team on the Tiangong 3.0 platform, tries to fix all three problems at once by fusing six cameras and a LiDAR into a single real-time 360-degree scene that a remote pilot or downstream policy can query from any angle.

The accompanying paper describes both the hardware rig and a feed-forward neural pipeline that turns sparse multi-camera input into compact, metric-scaled 3D Gaussians. The claim that got our attention: competitive rendering quality against recent generalizable Gaussian splatting methods while using roughly an order of magnitude fewer primitives.

Why forward-facing cameras keep failing

Many robotic systems rely on monocular or stereo cameras, or camera–LiDAR fusion, thanks to favorable trade-offs in cost, weight, power, and deployment complexity. Such configurations integrate onto diverse platforms and scale across environments. Teleoperators wearing head-mounted displays, however, hit three walls fast.

• A forward-facing camera's limited field of view restricts situational awareness. Operators cannot reliably perceive surrounding obstacles or affordances, which constrains safe motion and often forces uncomfortable body or head rotations to "search" for viewpoints.
• Unlike the human vestibulo-ocular system that stabilizes perception during locomotion, robot-induced jitter and motion blur destabilize the HMD view, causing simulator sickness and shortening effective operation windows.
• Multiple isolated camera feeds force context-switching, adding cognitive overhead during data collection or emergency takeover.

Panoramic fisheye rigs help, but they assume all lenses share (or nearly share) an optical center. RobotPan's cameras cannot be co-located due to hardware constraints, which makes classical image-stitching pipelines inapplicable and demands principled geometry-aware view synthesis. That is the geometric problem the system actually solves.