NeuraPick Robotics · Agri-tech · Edge AI

EVE-01 — autonomous harvesting at commercial speed.

A dual-arm robotic platform for greenhouse strawberry harvesting. It sees, reaches, picks and places — 1,200+ times an hour — with millimetre-accurate 3D vision and surgical-grade end-effectors. Edge-first: no cloud dependency, works where the WiFi doesn't reach.

1,200+
Picks per hour
3–4s
Pick cycle
IP65
Protection
48V
All-electric
01 Dual-arm architecture

Two arms. Six-hundred picks each. Every hour.

EVE-01 carries two independent manipulator arms, each with its own cable-driven, Da Vinci-inspired gripper with an integrated cutting blade. The arms work in parallel across the plant, doubling throughput without doubling footprint.

  • 600 picks per arm per hour — sustained, not peak
  • Da Vinci-style cable-driven grippers with integrated cutting blade
  • Redundant operation: arm B keeps working if arm A is mid-recovery
  • IP65 enclosure — runs in the humid, splashy greenhouse all day
02 Perception & vision

Millimetre-accurate 3D. Every frame, on device.

OAK 4D stereo cameras deliver dense 3D depth; a YOLO-based detector identifies ripe fruit by colour, size and occlusion state, and hands a list of candidate picks to the planner — all at 14 ms per frame on-device.

  • OAK 4D stereo cameras with on-board depth at 30 FPS
  • YOLO-family ripeness detection tuned on grower-specific datasets
  • Occlusion handling: plans around leaves and adjacent fruit
  • Self-calibration on start-up; no manual reprojection needed
03 Edge AI · local-first

All inference on the robot. Zero cloud required.

Every inference — perception, planning, control — runs on an NVIDIA Jetson module on the robot itself. No cloud dependency means no latency from the field, no outage when the farm's 4G drops, and no data-residency questions for growers.

  • NVIDIA Jetson edge compute — perception + planning local
  • ROS 2 native middleware with real-time QoS profiles
  • Offline-first fleet telemetry with resilient sync
  • Grower-controlled data — imagery never leaves the farm unless opted in
Specs

Built for the greenhouse floor.

Throughput
1,200+ / hr
600 per arm, sustained commercial speed.
Pick cycle
3–4s
Detection → plan → pick → place.
Harvest speed
0.1 m/s
Row-by-row scanning gait.
Transit speed
80 m/min
Between beds and charging stations.
Power
48V
All-electric, hot-swappable battery packs.
Payload
400 kg
Plus onboard crate rack.
Protection
IP65
Dust-tight, water-jet resistant.
Compute
NVIDIA Jetson
Edge inference — no cloud.
Middleware
ROS 2
Native navigation stack.
Stack

What it's running.

ROS 2 Python C++ NVIDIA Jetson YOLO v8 OAK 4D stereo OpenCV PyTorch Cable-driven actuators RadioMaster ELRS
← Previous project

Telemedicine

Have a project like this?

Start a project