Russian Lancet drones are not working as planned

A visual record dated January 29 illustrates an attack by a Lancet against a CV-90 infantry fighting vehicle, during which the “Target Locked” signaling is activated.

The Lancet, a kamikaze-type unmanned aerial vehicle developed by Russia, has emerged as one of the most outstanding instruments in the current theater of operations. This roving munition, noted for its affordability and weight of 35 pounds, has proven effective against a variety of targets, including artillery, armor and other strategic targets, reaching ranges of up to 40 miles.

Recently, the Lancet has undergone significant upgrades, incorporating new targeting technology that potentially marks it as the first fully autonomous combat weapons system, with the ability to identify and engage targets without human intervention. However, evidence suggests that its performance has not met anticipated expectations, leading to the possible deactivation of its automated mode.

Detailed analysis of failures in the Lancet autonomous system

Russian Lancet drones are not working as planned
Russian Lancet drones are not working as planned

Lancets frequently operate in conjunction with reconnaissance drones, also produced by ZALA, the company behind the Lancet. This collaboration allows reconnaissance drones to locate targets and transmit their coordinates to the Lancet operator, who then directs the drone toward the target for neutralization.

Despite this, developers have boasted about the advanced level of artificial intelligence built into the Lancet, and recent updates appear to give it even greater autonomy. In an interview with REN TV in December, Mikhail Kuzovkin, a military analyst, noted that the target acquisition process has been completely automated, eliminating the need for operator intervention.

“The Lancet now programs its mission independently, without requiring external targeting assignments,” Kuzovkin said, suggesting that this independence makes it immune to interference.

The newspaper Arguments and Facts reinforced this perspective in January, indicating that the role of the operator has been considerably minimized thanks to the integration of computer vision technologies, facilitating automatic target acquisition and raising the level of system autonomy.

Alexandr Zakharova, director general of ZALA, has rejected the label “artificial intelligence” to describe the Lancet’s capabilities, arguing in favor of a clear distinction towards advanced automation.

“This drone does not operate through artificial intelligence, but through specific algorithms and decision-making processes,” Zakharov explained in a televised interview in July 2023, outlining a pragmatic approach to developing unmanned combat technologies.

An examination of the downed Lancets revealed that they are equipped with the NVIDIA Jetson TX2 module, an edge device designed to provide “true AI computing” to small, low-power systems such as drones.

The TX2 and its revolutionary role in drone autonomy

Russian Lancet drones are not working as planned
Russian Lancet drones are not working as planned

The TX2 module, no larger than a credit card, is a fundamental pillar in developing cutting-edge neural networks and machine learning technologies. With its ability to execute 30 trillion operations per second thanks to a set of highly specialized GPUs, the TX2 has become the brains behind advanced drones like the Skydio 2. The latter has been described by experts as “astonishingly intelligent” due to its ability to navigate autonomously, follow subjects, and evade obstacles without human intervention.

Despite export restrictions imposed by the United States, which theoretically prevent its sale to Russia, the TX2 appears to be available at an affordable price on platforms such as Alibaba, suggesting that its acquisition does not represent a significant challenge for entities such as ZALA.

The main question lies in the TX2’s ability to perform autonomous reconnaissance of objects in complex terrain.

TX2 reconnaissance capabilities on the battlefield

The attacks carried out by the Lancet, documented through social networks, offer a clear view of the priority targets of these drones, with approximately half directed towards artillery and rocket launchers. This demonstrates the efficiency of the Lancets as counterbattery tools. In addition, the videos allow us to appreciate the operational perspective during the confrontation.

In recent records, it has been observed that moments before impact, the legend “Цель захвачена” (“Target locked”) appears on the screen, accompanied by a box that tracks the moving target. This mechanism is reminiscent of modern anti-tank systems like the Javelin, where the operator sets the target before launch.

However, this “target-locked” marker has only been identified in limited attacks against tanks and artillery, suggesting that harder-to-detect targets may still require human intervention.

The distinction between operational control and autonomy in goal identification remains ambiguous.

“The presence of a bounding box following the target indicates the use of artificial intelligence in target detection by the Lancet. However, the extent of this capability is still a matter of speculation,” commented Zak Kallenborn of the CSIS Strategic Technology Program. “The Jetson TX2 chip shows potential for basic object recognition, but determining the range and recognition accuracy of the Lancet is crucial.”

The TX2 could enable the ability to attack previously recognized targets. autonomously
“The basic autonomy to execute an attack once a specific type of target has been identified does not appear to require overly complex processing,” Kallenborn added.

Samuel Bendett, an expert on Russian drones and advisor to CNA and CNAS, noted that there are competing claims about the Lancet’s level of autonomy, ranging from semi-automatic to fully autonomous.

Russian Lancet drones destroy Ukrainian anti-aircraft systems in just one day.
“There are both versions, driven by the Ministry of Defense’s desire to promote it as a revolutionary system equipped with artificial intelligence,” Bennett explained.

Therefore, the Lancet could be interpreted more as an autonomous weapons system than one powered by artificial intelligence per se.

Challenges in targeting accuracy of autonomous systems

A visual record dated January 29 illustrates an attack by a Lancet against a CV-90 infantry fighting vehicle, during which the “Target Locked” signaling is activated. However, the Lancet, on its final approach toward CV-90, veers abruptly into a pile of debris at the last moment.

This incident, highlighted by the Military Informant account that shared the video, highlights a deficiency in the newly implemented target acquisition system, which mistakenly identified debris instead of an enemy combat vehicle, culminating in an operational failure.

This event highlights the inherent vulnerability of AI-powered computer vision systems, prone to errors in judgment that, to human eyes, are clearly wrong. Deep learning technologies can be confused by unusual perspectives on objects or easily fooled by things as insignificant as a sticker. An extreme example of this vulnerability was demonstrated in a study where an AI mistook a plastic turtle for a rifle.

It is plausible to consider that numerous similar errors are not publicly documented, given that such errors are rarely disseminated on social networks. Recent logs show a notable absence of the “Target Locked” indication and associated bounding box in Lancet attacks over the past two weeks.

This pattern suggests that the automatic target recognition software was hastily released and subsequently withdrawn for review. Previously, the system failed to identify complex, camouflaged targets, and currently, it fails even against simple, exposed targets, as evidenced by recent attacks on combat vehicles and tanks.

Lancet destroys a Ukrainian MiG-29.

Although it would be tempting to attribute the hit of a Lancet against a Russian T-90M to an AI failure, this incident, recorded before the software update, appears to be the result of traditional human error.

The implementation of automatic recognition in the Lancet does not yet define whether we are facing the advent of “Terminator-style robots” or if it is simply an auxiliary aiming tool. However, ZALA’s ambition to develop fully autonomous drone swarms remains evident while work is being done to fix bugs for a new attempt.

Samuel Bendett points out that other Russian groups are advancing in the development of neural network systems for small drones capable of identifying objects and equipment in the field with high precision, including tanks and combat vehicles. With technologies superior to the TX2, launched in 2019, the evolution of these systems is imminent.

Meanwhile, the debate on the regulation of autonomous weapons continues in international forums such as the UN, hoping to establish binding legal frameworks by 2026. However, by then, the use of these technologies may already be a consolidated reality on the battlefield.