New in the world of SLAM
We build spatial intelligence for robots. Right now that means a SLAM based drone, built and tested in simulation first, so it can map, plan and fly where GPS cannot reach.
What we work on
Four areas, one goal: a robot that knows where it is and decides well what to do next.
SLAM and path planning
Visual and inertial mapping that keeps a robot localized indoors, underground and anywhere satellites drop out, with planners that turn the map into a safe route.
Deep learning
Perception that runs on the robot itself: detection and recognition today, and learned 3D models that make maps denser and tracking more reliable.
Reinforcement learning
Robots that learn decisions from experience, built on the math of Markov decision processes.
Robotics and robot learning
Robots that learn skills by watching people, and the hardware and control that let them act on what they learn.
Veg, our drone that flies without GPS
Veg maps its surroundings with visual inertial SLAM, plans its own route, recognises people and objects onboard and lands safely after a rotor failure. The research version runs entirely on a Raspberry Pi 4 and an Arduino Nano, and is described in a paper on arXiv.
- Localization
- ORB-SLAM3, monocular and inertial
- Planning
- Dijkstra over a map built from SLAM points
- Control
- LQR attitude loop, PD position loop
- Safety
- Rotor fault detection with rerouted landing
- Vision
- Onboard object detection and face recognition
Now: building Veg 3.0 in simulation
Latest research
-
The "Faster Than Dijkstra" Algorithm Is Slower Than Dijkstra. The Benchmarks Prove It.
Headlines said a 2025 algorithm beat Dijkstra. On paper it does. In Python and C++ benchmarks, Dijkstra was 3 to 5 times faster. Here is why both are true.
-
Diffusion Policy Explained: How Robots Learn Multimodal Actions, with Python Code
Diffusion Policy explained from the Chi et al. paper: why behavior cloning averages actions, how denoising fixes it, the results, and tested Python code.
-
Beyond Imitation: How Reinforcement Learning Is Improving Robot Policies in 2026
Why 2026 robot policies are trained by imitation, then improved with reinforcement learning: π*0.6 Recap, RL-Co, FlowDPG and EXPO-FT, with Python code.
-
Robots Learning From Human Video: The Biggest Imitation Learning Shift of 2026
How robots learn from human video in 2026: EgoScale's scaling law, HumanScale, LAPT-Scale, HumanEgo and GR00T N1.7, explained with a small Python check.
-
DAgger Explained: How Dataset Aggregation Fixes Behavior Cloning, with Python Code
DAgger explained from the 2011 Ross, Gordon and Bagnell paper: why behavior cloning drifts, how dataset aggregation fixes it, the theory, and tested Python code.
Build with us
We are looking for research partners, pilot customers and engineers who want robots that work without GPS.