MIT researchers' UfM* cuts depth estimation uncertainty compute by 97%

Reliable uncertainty estimation is a critical bottleneck for deploying monocular depth neural networks in safety-critical robotics like autonomous drones or rovers. Traditional methods—ensembles and sampling-based techniques—require multiple forward passes per image, consuming excessive compute and memory. Worse, they often miss valuable signal from how predictions disagree across different viewpoints of the same scene.

To solve this, researchers from MIT (Soumya Sudhakar, Sertac Karaman, Vivienne Sze) introduce UfM* (Uncertainty from Motion*). Their key insight: represent the environment compactly with a Gaussian mixture and compute multiview disagreement efficiently between frames using only one DNN inference per image. Compared to a prior point–cloud approach, the Gaussian representation is far more memory-light and captures disagreement across continuous 3D regions. On 100 out–of–distribution ScanNet sequences, UfM* paired with aleatoric uncertainty slashes expected calibration error by 24–28% versus an ensemble, while needing just 3% of the energy and 0.02% of the memory. Deployed on an energy–constrained miniature robot with an Arm Cortex‑A76 CPU, the algorithm runs in real time at 30 FPS, consuming only 63 mJ per 224×224 image. This work enables robust, real‑time uncertainty for robots that must operate on a strict power budget, opening the door to safer autonomous navigation in the wild.