Adaptive AgroTechADAPTIVE AGROTECH
People behind the engineering

Maryam Behjati.

Making agricultural robotics visible, tangible and interactive.

RoleRobotics Engineer
CompanyAdaptive AgroTech
Joined2023

Maryam Behjati is a mechanical engineer with B.Sc. and M.Sc. degrees from the University of Tehran, where she received her master’s degree with highest honours. Since 2023, she has worked as a Robotics Engineer at Adaptive AgroTech, focusing on agricultural mobile robots, 3D modelling and dynamic simulation. Her work spans sensing, control algorithms, communication protocols and actuation, with a particular interest in affordable, small-scale robots for unstructured orchards and plant and soil monitoring. She models steering mechanisms and robot chassis, evaluates their controllability, reliability and manoeuvrability for agricultural work, and investigates modular sensors and the placement of NDVI and multispectral cameras, including the effects of bushes and occlusions on the information a robot can collect. She created Adaptive AgroTech’s interactive 3D demonstrations, translating engineering models into accessible experiences that help visitors explore agricultural robotics, sensing and automation. In January 2026, she received Adaptive AgroTech’s €2,000 Engineering Award in recognition of her significant contributions to the Adaptive AgroTech network and the development of its interactive simulation demo section, sharing agricultural engineering knowledge through models that visitors can explore directly.

Explore the navigation lab Publications Maryam’s 3D demonstrations ↗
Maryam@AdaptiveAgroTech.com Google Scholar Professional biography Engineering Award announcement ↗
Maryam Behjati, Robotics Engineer at Adaptive AgroTech
Maryam Behjati
Robotics Engineer

Shared experiment settings

Changing a numerical setting resets every robot. Camera and display layers do not change the simulation. All methods use a 0.32 m robot radius + 0.12 m safety margin, 0.75 m/s² acceleration and 1.8 rad/s² angular acceleration.

Range rays illustrate intersections with the known map; they do not supply sensor measurements to the controllers. Planner overlays show actual APF attraction/repulsion or the Pure Pursuit lookahead target. APF arrows use teal for attraction, amber for repulsion and violet for the resultant; their lengths are scaled and capped for display, not physical force units. Cinematic orbit moves all four cameras together.

Read the comparison correctly

Every panel shares the same flat pasture, stone positions, initial pose, goal, robot footprint, speed limit and 0.10 s simulation step. The default scene places three stones across the robot’s direct route. New stone layout generates another seeded arrangement for all four methods. Goal, stalled, no-route and 90 s time-limit outcomes are shown explicitly. Finished robots hold their pose while the shared scenario clock continues; any later contact is reported.

Travel time is simulated motion time. It excludes route precomputation and processor time. Compare travel time and path length only among successful runs. Minimum extra gap is a conservative swept-step clearance beyond the 0.44 m safety footprint and includes field edges.

The lower two methods have a global route; the upper two are local methods. A*/RRT prefer an additional 0.25 m route clearance and use the same tracking controller. These are deliberately different navigation pipelines, not a controlled benchmark of four interchangeable local controllers. Their parameter choices can change the ranking.

Make a video

Click Record video to reset all panels, record a 3-second lead-in, run them together and hold the final results for two seconds. Click Stop & save to finish early, then use the download link. The built-in export composes a clean 1920 × 1080 video with method labels, clock and live metrics; it records no microphone audio and asks for no screen-sharing permission. WebM is preferred; MP4 is used only if the browser supports it. Recording targets 30 fps; actual frame rate depends on your computer. Keep this tab visible. Camera, settings and stone-layout changes are locked during recording.

For external screen recording, use Fullscreen on a desktop display. The 3D pasture scenes, simulation scripts and profile portrait are embedded and work offline. Drag or scroll in any scene to adjust all four views together. If WebGL is unavailable, a matching top-down diagram preserves the simulation. On narrow screens the comparison board scrolls horizontally to preserve the 2 × 2 arrangement.

This is an educational simulation with exact obstacle geometry and a known crossing-rover trajectory and ideal planar motion. It does not establish real-robot performance or safety. APF principle: Khatib (1986). RRT principle: LaValle (1998). Farm rendering uses embedded Three.js under the MIT licence.

The pasture boundary is 13 × 6 m. Stones use the same circular collision models and inflated safety envelopes in each panel. The ground is flat: traction, grass resistance and terrain slip are not modelled.

Dynamic challenge and live graphics

In Crossing rover, the same scripted rover moves on the shared simulation clock. The temporal safety check considers its predicted motion during candidate trajectories and braking. A* and RRT retain their initial-map routes; their tracking safety filters may slow or wait for the crossing rover. This is not dynamic global replanning. Actual swept contacts are reported as a collision outcome. Speed traces show recorded simulation states. Goal proximity uses straight-line distance to the target and can decrease during a detour; completed runs show 100%.