From drive-by-wire hardware to embodied intelligence, every layer is available independently. Chassis hardware (L0) is manufactured by RJ Tech; L1–L4 and the data loop are developed in-house by TerraAgent Robotics.
This is the only layer that spans both brands. The multi-motor distributed drive, four-wheel independent suspension, and large-battery range-extended EREV chassis are manufactured and delivered by RJ Tech. The complete drive-by-wire interface for steering, braking and drive (documented CAN FD / ROS2 APIs) is defined and integrated by TerraAgent Robotics — much like the open-interface story RJ Tech previously built around SteerAI and Autel, now internalized as an interface agreement between the two brands. L0 hardware determines how fast L1–L4 can go, how steep they can climb, and how extreme an environment they can keep working in.
Distributed drive control (torque vectoring): millisecond-level independent torque allocation across four wheels for zero-radius turns and 35° hill climbs. Terrain-adaptive control: MPC/MPPI model-predictive control with PID as a safety fallback, staying stable under split-µ conditions (different tire grip left vs. right). RL injection ①: reinforcement-learning-trained control policies — terrain-adaptation strategies are explored and transferred in simulation (sim-to-real). Don't confuse this with L0's distributed drive: L0 is the multi-motor hardware; L1 is the control algorithm that coordinates those motors.
BEV multimodal fusion perception: camera + LiDAR + millimeter-wave radar fused into a unified bird's-eye view, outputting a traversability cost map (occupancy · slope · roughness · semantic layers) and an elevation map. Fused localization: RTK-GNSS + visual-inertial odometry (VIO) + LiDAR-inertial odometry (LIO) + SLAM, tightly coupled across all four sources. In GNSS-denied environments (tunnels, canopy, mines), LiDAR SLAM takes over seamlessly, holding centimeter-level accuracy. L2 + L3 together are what we call the Autonomous Driving stack — TerraAgent Robotics' core competency layer. The BEV / Transformer approach matured by the autonomous driving industry — rebuilt for terrain, re-targeted from driving to working.
One quick-release gantry carries the vehicle's eyes and brain: the OT128 hangs inverted, center-mounted, sweeping 360°; the EO gimbal takes the highest point; dual GNSS antennas set left and right give you heading; edge computing and dual comms all pack into the gantry. Four latches — attach it and you have an autonomous platform; release it and you have a great vehicle to drive. Most importantly: not an inch of cargo space or riding area is taken — it all stays for the mission. The same architecture runs on both the Big ATV and the Full-Stack UGV: the UGV uses it to stack redundancy and push the perception ceiling; the Big ATV uses it for embedded, centralized perception.
Perception shouldn't cover the whole vehicle. It only needs the rear thirty centimeters — the rest is yours.
The same interface also accepts: IR thermography, spray modules, searchlights, loudspeaker deterrents, weather stations, comms relays, drone docks — change the mission, change the payload.
The gantry blocks neither the seat nor the handlebar — the person rides, the vehicle keeps watch: you drive up front, it scans the whole terrain behind you.
Global path planning + local obstacle-avoidance replanning: safe, efficient driving paths generated in real time on top of L2's cost map. Path memory (teach & repeat): the operator drives the route once, the system remembers it and replays it automatically — the baseline capability for rapid field deployment. Follow-Me: the vehicle follows the operator autonomously. Fleet coordination: multi-vehicle task allocation, path coordination, formation dispatch — proven at the 100-vehicle scale.
L4 sits on top of the Autonomous Driving stack (L2+L3) and is independent of it — the arm only comes after the vehicle moves. Dual-arm autonomous operation: vision-language-action (VLA) / vision-action (VA) models drive cooperative arms through delicate field work (PV panel replacement, valve operation, sample collection). Air-ground coordination (arm × drone): the arm launches, catches, docks and recharges drones. Physical AI: transferring manipulation policies from simulation to the real world. RL injection ②: sim-to-real training of manipulation policies. Technician AI is the direction we're building toward — an off-road chassis with 1–2 lightweight arms, as a transitional step. This capability is still ahead of us; we're building the vehicle first. KUKA arm integration is already validated.
This isn't a standalone product layer — it's a thread running across L1–L4. Data loop: collection (every delivered Full-Stack UGV collects real off-road scenario data) → annotation (LiDAR generates dense depth ground truth for cameras) → cleaning → re-feed (feeding the algorithm models of the Big ATV and Small ATV). Simulation: Isaac Lab / MuJoCo / Cosmos multi-platform simulation environments. RL training: reinforcement learning injected into L1 (control policies) and L4 (manipulation policies). Multi-vehicle VAL (Vehicle Abstraction Layer): one autonomy stack, adapted through the vehicle abstraction layer to three different bodies — the Full-Stack UGV, Big ATV and Small ATV.
The first six are platform baseline capabilities — every TerraAgent platform has them; the second six are the core autonomous driving capabilities, spanning L1–L3 and forming the technical backbone of the autonomy stack.
You drive the route once; it remembers it once. From then on, that route is its job.
Mount anything on the deck — cargo boxes, tanks, equipment — swap it out by the day's task.
A controller can drive it; a tablet can too. You watch from afar while it works on site.
Mud, slopes, rubble — all fine. Where road vehicles stop is exactly where it clocks in.
It shuttles the heavy loads around the site, so people don't make the trips.
Built to a tough standard — sun, wind and rough use are all in a day's work. It's made for the rough stuff.
Camera, LiDAR and millimeter-wave radar fused into a single occupancy grid — every cell priced: drivable or not, and how hard.
RTK + visual-inertial + wheel speed; when signals drop, SLAM takes over — centimeter-level accuracy holds under tunnels and canopy.
Spots an obstacle and reroutes on the spot; slows down early on rough roads. No heroics — discretion is a virtue.
MPC in charge, PID as fallback, each wheel pulling its weight. The slicker it gets, the more stable it stays.
Follow, remote, full autonomy — switch freely between the three. You walk; it follows — a pack mule that never asks for feed.
One control center manages a whole fleet — the order is clear, positions aligned to sub-meter accuracy.
From the L0 drive-by-wire interface to L4 embodied intelligence, we have a detailed technical whitepaper for every layer.
Contact us · request the technical docs