How Robot Dogs See the Unseeable: Improving Visual Interpretability via Peering for Exploratory Robots

If our eyes would be 1m in diameter, we would not need to do this.

Prior to peering, the robot perceives only the occluding foreground vegetation, limiting vision reasoning to the immediate surface of that vegetation. After peering and synthetic aperture processing, the robot reveals the entire scene behind the vegetation across a range of depths (within physical limits), enabling vision reasoning models to successfully interpret the background environment.

First, robotic platforms must be capable of navigating highly complex environments. Autonomous navigation relies directly on scene understanding, which in turn requires robust sensing despite severe vegetation-induced occlusion. This requirement is fundamental. While specific applications demand different robotic platforms, both the underlying challenge and its solution remain consistent. For instance, drones are often more efficient than ground robots for search-and-rescue operations, yet they face similar top-down canopy occlusion. Conversely, ground robotics applications in unconstrained outdoor environments often involve defense or tactical scenarios. Beyond general occlusion removal, our paper explores de-camouflage as a key application - evaluated first in simulation and subsequently on physical robotic platforms. The underlying principle is identical: while well-camouflaged targets evade detection by human observers and standard vision-reasoning systems, they can be revealed when leveraging motion parallax, a mechanism inspired by animal behavior. Broadly, this approach equips exploratory robots with robust scene-understanding capabilities in complex, vegetated environments.

Earlier this year, I participated in an expedition to locate the last remaining wild Sumatran rhinos in the dense rainforests of Indonesia. The species had not been sighted for nearly a decade, despite extensive monitoring efforts using drones, camera traps, eDNA sampling, and trained tracking dogs. Extreme humidity and dense jungle terrain severely constrain tracking dogs, suggesting that quadrupedal field robots equipped with chemical or olfactory sensors could eventually offer a viable alternative - provided they can navigate such dense environments. Ultimately, using our drone platforms, our expedition successfully located and verified a wild Sumatran rhino in just five days.