All projects
Completed 2023 – 2025 Strategic Data Solutions Lab, Hitachi America Ltd.

Unmanned Aerial Vehicle Safe Autonomous Operations

Safety guarantees for autonomous drones and air taxis: learned reachable sets that build collision avoidance into the controller, and camera-based 3D models from neural radiance fields for safe path planning.

Collaborators: Malarvizhi Sankaranarayanasamy, Ravigopal Vennelakanti

Paper Patent Slides
Two animations side by side: a 2D slice of the value function for two UAVs avoiding a collision, evolving over time with its zero level set in black, and the matching reachable tube rendered as a rotating 3D volume
Two UAVs avoiding a collision, a standard benchmark known as Air3D. Left: a 2D slice of the value function as it evolves, with the zero level set (black) marking the boundary of the states that can lead to a collision. Right: the full reachable tube in 3D, over relative position and heading. From Embedding Safety Requirements into Learning-Based Controllers for Urban Air Mobility Applications, AIAA SciTech 2024.

Problem

Air taxis and delivery drones are expected to fly autonomously over cities, but public trust is low: in one survey cited in this work, only 26% of respondents trusted autonomous air taxis. Before these vehicles can fly on their own, their controllers need a clear safety guarantee: the set of states from which the vehicle can still avoid a collision or reach its landing pad. Computing that set means solving a Hamilton-Jacobi equation, and grid-based solvers become impractical for realistic vehicle models with 6 to 12 states. Learning-based controllers scale better, but they are hard to certify.

Approach

The first line of work learns the safety guarantee instead of gridding it. A physics-informed neural network is trained to satisfy the Hamilton-Jacobi equation and its boundary conditions, building on the open-source DeepReach framework. The learned value function gives the reachable set directly, and the safe control follows from it. Results are checked against hj_reachability, a JAX-based numerical solver.

The second line of work brings camera-based 3D models into path planning. A neural radiance field, trained with nerfstudio on images from several viewpoints, turns an object into a 3D point cloud. Its convex hull can stand for the robot’s own shape, an obstacle or the goal. The robot’s reachable sets are computed in closed form as polytopes, so a model predictive controller only has to satisfy linear inequality constraints to stay clear of every obstacle.

Four panels: point clouds of an object recovered from a neural radiance field at four viewing angles, their convex hull, a planned path through a field of box obstacles, and a close-up of the convex hull at the goal

From images to a safe path: (1) an object recovered from a neural radiance field, seen from several angles, (2) wrapped in a convex hull, (3) a collision-free path through randomly placed obstacles to a goal defined by that object, and (4) a close-up of the goal. Adapted from Safe Navigation using Neural Radiance Fields via Reachable Sets, ICMCR 2026.

Outcome

  • Learned safety that matches a numerical solver. On the Air3D benchmark, where two UAVs must avoid a collision, the physics-informed network reproduces the nonconvex reachable tube computed by a numerical solver.
  • Safe vertiport landing. The same approach finds every state from which a UAV can land safely on a vertiport.
  • Collision-free planning with NeRF models. Closed-form reachable sets keep the planner clear of obstacles in two cluttered scenarios, using a NeRF object as the robot’s shape in one and as the goal in the other. The NeRF model trains in about 18 minutes on a single CPU.
  • Published and patented. Papers at AIAA SciTech 2024 and ICMCR 2026, and a patent application on AI-based safe control for unmanned aerial systems.