Stanford Robotics Seminar ENGR319 | Winter 2026 | Resilient Autonomy

Underground exploration demands full autonomy

Systems must operate with intermittent or no communication to base stations due to rock interference, requiring fully on-board mapping, planning, and object detection in caves, coal mines, and decommissioned nuclear plants.

Robustness over peak performance

Unlike current research optimizing for brittle high-performance metrics, resilient autonomy prioritizes robustness through redundancy and resourcefulness to handle dust, darkness, and featureless corridors without system failure.

Real-world validation in Pittsburgh

Testing occurred in diverse challenging locations including a former veterans hospital, limestone mines, and a coal mine cabin to ensure algorithms function across big open spaces, narrow passages, and urban structures.

Single encoder replaces task-specific models

Map Anything processes pixels once to perform multiple tasks—monocular depth, SFM, depth completion, and localization—avoiding GPU bottlenecks caused by running separate object detectors, SLAM, and VQA models simultaneously.

Factorized outputs enable arbitrary cameras

Unlike prior work (DUSt3R, VGGT) requiring redundant prediction heads, this architecture uses a single DPT head predicting scale and ray directions, allowing flexible inputs for calibrated or uncalibrated cameras and auxiliary data like poses.

Superior odometry with monocular input

Compared to off-the-shelf stereo visual-inertial odometry (ZX camera baseline), Map Anything achieves significantly less drift running at 15-16 Hz using only monocular images, and upcoming optimizations promise 7-10x speedups.

Thermal alignment with visual backbones

AnyThermal fine-tunes DINOv2 on thermal data to align thermal and visual representations, enabling existing visual pipelines (segmentation, place recognition) to work natively with longwave infrared for night and dust penetration.

Open-source synchronized dataset collection

The TartanRGBT dataset uses a custom open-source platform with hard-synchronized stereo thermal/visual sensors, an AGX, and a physical button (no typing in freezing field conditions) to address the scarcity of quality thermal training data.

Online IMU learning for vision degradation

When LiDAR or vision fails in featureless hallways or complete darkness, the system refines a pre-trained IMU model on-the-fly using available visual data, then relies solely on the IMU (which cannot be jammed) to maintain accurate odometry.

Native spherical convolutions for fisheye lenses

Rather than projecting distorted fisheye images into multiple virtual pinholes, the system projects images into a canonical spherical space with specialized convolution and pooling, enabling efficient 360-degree perception for drones without edge distortion.

Unified Flow Matching for correspondence

UFM combines wide baseline matching and optical flow using co-visibility masks to handle large viewpoint changes or small motions, providing dense pixel correspondence critical for manipulation and localization tasks.

Ray Frustons for long-range semantic reasoning

This representation combines 3D semantic voxel maps with directional ray vectors at frontier points, allowing the robot to navigate toward distant semantic targets (like water towers) visible in images but beyond accurate depth sensing range.