Desert Robotics develops embodied intelligence, autonomous systems, and advanced robotic control for machines operating beyond controlled environments.
Robots are leaving controlled environments. The next generation must perceive, learn, adapt — and act safely in the physical world. AI learned to understand the digital world. We build the systems intelligence needs to act in the physical one.
Robots that understand instructions, perceive their surroundings, and acquire reusable physical skills.
Intelligent machines capable of operating in unstructured and uncertain environments.
The layer that turns machine intelligence into safe, precise physical behavior.
A mobile manipulation platform executing natural-language tasks through vision-language-action policies, from teleoperated demonstration collection to autonomous deployment.
Reinforcement-learned agile flight through cluttered environments, trained in simulation with domain randomization and transferred toward real quadrotor hardware.
Control policies and learned action spaces defined on the geometry of physical configuration spaces — where robot states live on manifolds, not flat vectors.
Why geometry matters. A robot's orientation lives on SO(3), a curved manifold — not a flat vector space. Play both interpolations between the same two poses: the geodesic stays a valid rotation at every instant, while naïve Euclidean averaging of matrix entries shears and collapses the frame. Our research asks how learned policies should output actions Δq ∈ se(3) that respect this structure.
↑ FEEDBACK — VISION · PROPRIOCEPTION · FORCE · TACTILE ↑
How can learned robot behaviors transfer across tasks, environments, and robot morphologies — instead of being retrained from scratch for every machine?
How should robot foundation models produce actions when configuration spaces are manifolds rather than flat Euclidean vectors?
How can robots learn compliant interaction, force regulation, and impedance behavior — including when rich force sensing is available only during part of training?
How can autonomous machines remain predictable and recover safely under uncertainty, model errors, sensor failures, and adversarial conditions?
| Robot Intelligence | Vision-language-action models, VLMs, multimodal agents |
| Robot Learning | Reinforcement learning, imitation learning, offline learning |
| Control | Geometric, nonlinear, force and impedance control |
| Autonomy | Navigation, planning, perception |
| Simulation | MuJoCo, PyBullet, Gazebo, Isaac Sim |
| Robotics Infrastructure | ROS 2, real-time middleware, edge deployment |
| Hardware | Embedded systems, sensing, rapid prototyping |
| Security | Robotic and cyber-physical systems security |
Most learned policies emit actions as flat vectors. But the spaces robots actually move in are curved. What breaks when we ignore that — and what a geometry-native action head looks like.
Read note →From teleoperation rig to demonstration dataset to first autonomous rollouts: the unglamorous engineering behind a language-conditioned robot, and what actually failed along the way.
Read note →Domain randomization, latency modeling, and the disturbance-rejection layer that decides whether a learned flight policy survives contact with real air.
Read note →Ibrahim Al-Shehri is a robotics researcher and engineer working at the intersection of autonomous systems, geometric control, and embodied artificial intelligence. His background spans electrical engineering, physics, mathematics, AI, and cyber-physical security — a convergence the company is built on.