Autonomous payload-drop drone for NIDAR, AeroTHON and similar contests

Most Indian drone contests come down to the same mission: search an area, find a marked target with a camera, put a payload on it and come home, all without a pilot. This is the hardware and the set-up for that mission, and the sums that decide whether the payload lands where you meant it to.

The mission these contests share

NIDAR, the National Innovation Challenge for Drone Application and Research, is run by MeitY and the Drone Federation of India under the SwaYaan programme. Student teams build autonomous drones for image-based detection and autonomous delivery, with disaster management and precision agriculture as the themes. NIDAR 2.0 runs through 2026–27.

SAE AeroTHON 2026, in its rotorcraft track, asks for a fully autonomous aircraft that finds the correct location, lowers a payload to the target while holding its mission altitude, and returns to the take-off point through a corridor while avoiding obstacles.

The details change every year, so read the current rulebook before you buy anything. Weight limits, payload size, geofence and kill-switch rules shape the design more than any part choice. Under the details, both contests ask for four things:

1. Fly a search pattern on its own.
2. Recognise the target in a camera image and work out where it is on the ground.
3. Put the payload on it, by dropping it or lowering it.
4. Come home safely, whatever went wrong on the way.

The system

• The autopilot flies everything: the search pattern, the approach, the return and every failsafe. We use ArduPilot on a Pixhawk 6C here. PX4 works as well if your team already knows it.
• A companion computer reads the camera, runs the detector and tells the autopilot where the target is over MAVLink.
• A release or winch hangs off one of the autopilot's servo outputs, so the drop is a mission command. Your Python code never drives it directly.
• A ground station on a telemetry link, and an RC transmitter so the safety pilot can take over at any moment.

Airframe: size it from the payload

Start with the all-up mass: airframe, battery, companion computer, camera, the release mechanism and the heaviest payload the rules allow. Then check that the motors give at least twice that in full-throttle thrust, and estimate the flight time. Our final-year project guide works through both sums with real numbers.

For this mission we suggest the Holybro X650. Its 650 mm folding carbon frame carries T-Motor MN4014 KV330 motors on 15-inch propellers, a propulsion set built to lift rather than to race, and the centre plate has room underneath for a companion computer, a downward camera and a release. Holybro recommends a 6S LiPo of 5,000 to 12,000 mAh. The development kit comes with a Pixhawk 6C, an M10 GPS, a PM02 V3 power module and a SiK telemetry radio. The smaller X500 V2 is cheaper, but a payload, a release and a Jetson quickly use up its margin. Our kit comparison sets the three side by side.

Finding the target

Choose the altitude from the pixels you need. A detector needs the target to cover some minimum number of pixels, commonly 20 to 30 across. The ground sample distance, the width of ground one pixel covers, is:

GSD = sensor width × altitude ÷ (focal length × image width in pixels)

Take the Raspberry Pi High Quality Camera, whose sensor is 6.29 mm wide (4,056 pixels of 1.55 µm), with a 6 mm lens, flying at 20 m. The camera sees a strip 6.29 ÷ 6 × 20 = 21 m wide. The detector will usually run on a reduced image, say 640 pixels wide, so each pixel covers 21 m ÷ 640 = 3.3 cm. A target 1 m across then spans about 30 pixels, which is enough. Fly higher and the survey is quicker but the target shrinks. Space the lanes of the search pattern at about 80% of the strip width, so neighbouring images overlap.

Train on your own pictures. A small YOLO-type model trained on photographs of the actual target, taken from the altitude you will fly, beats a large model trained on anything else. A Jetson Orin Nano runs such a model at useful frame rates. A Raspberry Pi 5 manages a few frames per second, which is enough for a slow search.

Point the camera straight down on a vibration-damped mount, and calibrate it with a checkerboard in OpenCV so you know its focal length and distortion.

Where is it? The accuracy sums

When the detector finds the target, the companion computer turns a pixel into a ground position from the aircraft's position, altitude and attitude at the moment the frame was taken. Each of those carries an error:

• GPS. A standard receiver such as the M10 is typically good to a metre or two.
• Attitude. An error of 2° in roll or pitch at 20 m moves the answer by 20 × tan 2° = 0.7 m.
• Timing. At 5 m/s, pairing the image with a position 100 ms out of date puts the target 0.5 m off.

Together they can put the target two or three metres off, which may miss a small drop zone. There are two ways to fix that.

The cheap one is visual re-centring. Once the target is found, switch to GUIDED, fly over the estimated position, descend, and nudge the aircraft until the target sits in the middle of the image. The GPS and timing errors then no longer matter, because the camera is measuring the target directly. Build this first.

The other is RTK GNSS, which brings position error down to centimetres but needs a base station and a correction link. Our GPS versus RTK guide covers what it takes.

Dropping versus lowering

A drop uses a servo release on one of the autopilot's servo outputs. In a mission it is a DO_GRIPPER or DO_SET_SERVO command after the waypoint over the target. The companion computer can send the same command over MAVLink. A dropped payload keeps the aircraft's forward speed, and it falls from 10 m in √(2 × 10 ÷ 9.81) = 1.4 seconds. Released at 2 m/s it would land about 3 m ahead, so always release from a steady hover.

Lowering needs a winch. ArduPilot drives a servo-based winch from a mission with the DO_WINCH command. A payload on a line swings, so lower slowly and hold position while you do it. If the rules let the aircraft descend instead, ArduPilot's NAV_PAYLOAD_PLACE command descends until it senses touchdown and then opens the gripper.

Putting the mission together

1. Search in AUTO. Plan a lawnmower survey over the search area in Mission Planner or QGroundControl, at the altitude you worked out above.
2. Detect. The companion computer watches every frame and logs each detection with the aircraft position that goes with it.
3. Approach in GUIDED. On a confirmed detection, the companion computer switches the autopilot to GUIDED and sends position targets over the estimated location.
4. Re-centre and deliver. Descend to the release height, centre the target in the image, hold, then send the release or winch command.
5. Return. Climb and switch back to AUTO for the return leg through the corridor, or to RTL if the rules allow a direct return.

Keep the safety logic in the autopilot. The geofence, the RC-loss and telemetry-loss failsafes, the battery failsafe and a motor emergency-stop switch on the transmitter all work there even if the companion computer's code crashes. Contest scrutineers check these before anything else.

Companion computer options

The simplest arrangement is a separate computer, a Jetson Orin Nano or a Raspberry Pi 5, on the Pixhawk 6C's TELEM2 port. Our companion computer guide covers the power supply and the bench tests. If you want the autopilot and the companion on one board, Holybro's Pixhawk Jetson Baseboard takes a Pixhawk 6X module and a Jetson Orin Nano or Orin NX, linked by UART, CAN and Ethernet. That is a neater build, but more to buy and more to learn.

Before you fly in India

With a battery this aircraft is likely to weigh more than 2 kg, which makes it a small drone under the Drone Rules, 2021. Rule 42 exempts research and testing flights by educational institutions recognised by the central government from the type certificate, unique identification number and remote pilot certificate when they fly within their premises or a green zone. At the contest, follow the organisers' instructions. For practice flights off campus, check the Digital Sky airspace map and ask your department first.

Sources.

• NIDAR's organisers, themes and 2026–27 edition from MyGov and nidar.org.in (accessed 2026-09-23)
• AeroTHON 2026 rotorcraft mission from the Track 1 rulebook (accessed 2026-09-23)
• X650 kit contents and recommended battery from Holybro's X650 development kit page (accessed 2026-09-23)
• Gripper, DO_SET_SERVO and NAV_PAYLOAD_PLACE from ArduPilot's servo gripper page and mission command list (accessed 2026-09-23)
• Winch control from ArduPilot: control a winch (accessed 2026-09-23)
• Sensor resolution and pixel size of the High Quality Camera from the Raspberry Pi product brief (accessed 2026-09-23)
• The R&D exemption and drone categories from the Drone Rules, 2021 backgrounder published by PIB.

Published 23 September 2026. Last revised 25 September 2026. Corrections to sales@vebixautomation.com.

Entering a contest?

Send us the rulebook's weight and payload limits and your timeline. We will check the parts against them and send a quotation your college can raise a purchase order against. Tell us the submission date, and we will tell you honestly whether the delivery can meet it. Ask for a quote.