An existing mechanical machine becomes an edge-AI controlled autonomous vehicle without structural redesign. The kit provides servo actuators, traction drive, field I/O and a vehicle control unit built on the NVIDIA AGX platform.
The machine's original controls are actuated, not removed and re-engineered. Every stage is a working machine — no stage is a prototype.
VCU, auxiliary node and gateway go into the existing electrical bay, powered from the host 12 V bus. Four servo actuators replace the deck-lift, governor and blade-clutch linkages. The gateway taps the engine CAN, so rpm, temperature and hour meter reach telemetry unchanged.
Result: beyond-line-of-sight operation from a portable panel, with recorded sessions as training data.
Hydrostatic drive out, two 20 kW induction motors onto the existing sprocket flanges. The 48 V pack and BMS go into the front bay. Traction becomes a torque command — slope-aware traction control and precise skid-steer.
Result: one operator supervises several machines from a single panel. The safety loop stays hard-wired.
A semi-solid-state 3D lidar on the front mast covers 360° to 90 m at 0.2° resolution — obstacle field, boundary and terrain classification. Four GMSL2 cameras add 120° front and rear and 90° each side, at 1928 × 1208 and 30 fps, over a 14 Gbps board-to-board link.
Slope-tolerant ground segmentation and foliage filtering reduce it all to a single obstacle cloud. Neural networks run hardware-accelerated on the machine itself.
| 3D lidar | 360° × 90 m · 0.2° |
|---|---|
| Camera front / rear | 120° · 1928 × 1208 @ 30 fps |
| Cameras left / right | 90° each |
| Camera link | 4 × GMSL2 · 14 Gbps |
The vehicle control unit carries perception, planning, telemetry and an independent safety co-processor in one sealed box. All field I/O lands on a single 48-pin connector.
| Compute | Jetson AGX Orin 64 GB — 2048 CUDA, 64 Tensor, 2 × NVDLA · 275 TOPS |
|---|---|
| Motion bus | CAN-FD · CiA 402 · 8 Mb/s — deck lift, governor, blade clutch |
| Implement bus | CAN-FD, opto-isolated — host engine ECU, traction inverters |
| Network | Six-port managed GbE — lidar, uplink, 5G, service port |
| Safety plane | Dual E-stop, radio-loss and tilt trip → traction disable. Hard-wired |
| Root of trust | TPM 2.0, secure boot, separate watchdog processor |
The machine follows a person, avoids obstacles and executes waypoint missions. It takes commands from hand gestures, a browser cockpit, or natural language by voice and chat.
Safety is structural, not conversational: motion starts only from a held idle state, and stop overrides everything. All motion converges into one mode-selected stream.
| Live cockpit | Full runtime graph in any browser, sub-second video |
|---|---|
| Gesture control | 23 classes, dual-hand, sliding-window vote |
| Operator identity | Hi-vis vest segmentation with tracked IDs |
| Following | Camera–lidar fusion at 10 Hz or better |
| Missions | Walk-and-record waypoints, heading-gated |