A robot shooting a recurve bow: Christmann 2022

A robotics group in Taipei taught a humanoid robot to grip, draw and release a recurve bow, and measured where the arrows landed. The robot is a THORMANG3, 137.5 centimetres tall, 42 kilograms, 29 degrees of freedom, with a camera in its head; the bow is a 66 inch Cartel Sirius Plus, 16 pounds at a 28 inch draw, its grip wrapped in foam so the gripper could hold it. The target is recognised by looking for three concentric circles with the Hough transform.

There are three experiments. In the first, the robot shoots ten times at a wall 10 metres away with no aiming correction, using a rubber arrow with a suction cup and a 65 centimetre draw: the two furthest impacts are 62.14 centimetres apart horizontally and 9.34 vertically, because after every shot the robot wobbled and turned slightly to the right.

In the second, from the same position, the robot aims with the camera, correcting yaw proportionally and holding roll fixed. The target is a 49 centimetre disc with a 9.7 centimetre inner circle, hung with its centre 114.5 centimetres above the ground. The ten impacts fall within 6.6 centimetres horizontally and 4.7 vertically, but the centre of the group sits 27.85 and 25.96 centimetres from the centre of the target: a tight group, a quarter of a metre off.

In the third experiment, in an open field with 32 inch arrows, the robot sent the arrow 44 metres at a 70 centimetre draw and 4 degrees of roll, 55 metres at 70 centimetres and 10 degrees, and 50 metres at 65 centimetres and 12 degrees. The distances were estimated by filming the shot and locating the landing point on satellite imagery.

A person nocks the arrow by hand before every shot: the robot does the rest. The left gripper holds the bow, the right one takes the string and releases it by opening. Both arms are driven by inverse kinematics, computed with the pseudoinverse of the Jacobian, and the drawing motion follows the vector from the arrow rest to the nock, because departing from it sends the arrow off target or knocks it off the bow. The string at rest sits 22 centimetres from the rest; from there the right gripper closes on the string and takes it to the required draw length.

Aiming works in two steps. First the camera finds the target: greyscale image, Gaussian blur with a kernel of 5, Hough transform with radii between 10 and 30 pixels at 10 metres, then eight profiles of pixel differences running outwards from the centre, non-maximum suppression, and a circle is accepted when three local maxima follow one another at consistent radii. At least five positive profiles are needed to declare the target found.

Then the correction, calibrated by hand. The robot shoots a few arrows with yaw at zero, the target is placed at the centre of the resulting spread and that position becomes the reference; the target is then moved sideways and yaw values that hit it are found by trial, giving the proportional gain. Roll stays at a fixed value.

In the first two experiments the robot was neither moved nor realigned, and the second was run straight after the first, with the target already fixed to the wall.

This is an arXiv preprint, called preliminary by the authors themselves, and it has not been through peer review. The material is a single robot, ten shots per condition in the two accuracy experiments and one shot per setting in the distance test.

In the two accuracy experiments the arrow is not an arrow: it is a rubber shaft with a suction cup, chosen for safety. The draw weight is 16 pounds, a fraction of what is used in competition, and the authors themselves note that reaching 60 or 70 metres would need stronger limbs. No shot was scored and no archer shot for comparison, so the parallel with human shooters remains a prospect rather than a result.

The systematic offset in the second experiment is not discussed. The ten shots fall within a few centimetres, but the centre of the group sits about 28 and 26 centimetres from the centre of the target: the calibration was gathering the group without bringing it to the middle, and the paper does not explain why.

The numbers in the first experiment do not hold together: the "variation" of 297.75 is given in centimetres but is clearly a variance, so the units in the results are inconsistent. In the third experiment distances are read off Google Earth from a video, with no data on wind or conditions, and the arrow is described as carbon fibre in the methods and as fibre glass in the results.

The study is not about technique, it is about machines, and for that reason it is useful for one thing: showing how much the mount holding the bow weighs on the result. With no aiming correction the ten shots opened up by 62 centimetres horizontally at 10 metres, while vertically they stayed under 10. That is not the bow: it is the robot shifting and turning to the right between shots, and the dispersion follows the direction of that rotation. With camera-guided aiming the ten shots came back inside roughly 7 centimetres by 5.

The other thing the numbers show is the difference between a tight group and a centred one. After calibration the robot repeated the shot within a few centimetres, but the group stayed a quarter of a metre off centre: repeatability of the movement and calibration of the sight are two separate problems, and solving the first does not solve the second.

None of this says what a bow does in a person's hands. The draw weight is 16 pounds, the arrow in the accuracy tests is rubber with a suction cup, there is no score and no archer in the comparison. The authors themselves frame the work as a first step towards a robot that could shoot at competition distances.