A 328-study review shows robot arm gains coming from control software and vision

Researchers pulled 1,246 records from five research databases, and 328 made it through screening. What those 328 studies measure tells a plant engineer where to look hardest when an AI-driven robot cell comes up for approval.

Ahmed Sameh, Mohamed Fanni and Maher Rashad led the systematic review, which Bioengineer.org summarized on November 11. It covers work published between 2010 and 2025. The team searched Scopus, Web of Science, IEEE Xplore, SpringerLink and ScienceDirect, screened the results under the PRISMA method, and kept peer-reviewed journal articles, conference proceedings and official industrial standards. Everything it found falls into four layers: mechanical, computational, perceptual and regulatory.

The authors start from the old model, where an arm sat behind a fence and ran a fixed program for spot welding, material transfer or fast pick-and-place. Plants now want arms that adapt on the line and work close to people. The review's evidence suggests that the stack running the arm decides most of whether it can do that. This fall, several large automation suppliers paid to own pieces of that stack.

The trade-offs built into the hardware

The review treats structure and actuation as the baseline every controller has to work within. The choices it describes are physical trade-offs, and a spec sheet should state them up front:

• Serial arms, such as standard six-axis articulated robots: large workspace and versatile dexterity, but positioning errors build up joint by joint.
• Parallel arms, such as Delta robots: high rigidity and micro-scale precision from closed structural loops, confined to a compact working volume.
• Hybrid designs: pair serial reach with parallel stiffness.
• Kinematically redundant arms: extra degrees of freedom that help avoid singularities and steer around obstacles.
• Electric servomotors: the dominant actuator, chosen for precision, easy electronic integration and compact size.
• Hydraulics: still used for heavy force, with fluid maintenance and containment to manage.
• Pneumatics: rugged and cheap for high-frequency sorting.
• Variable stiffness actuators and soft elastomer robotics: emerging options that shape contact forces to make close work with people safer.

For anyone buying reach-heavy cells, the serial-arm entry matters most. Errors that stack across joints are the kind kinematic calibration is meant to correct, and one of the strongest software results in the review targets calibration directly.

Where the measured gains turn up

Across research published from 2010 to 2025, the arm sets the physics, and the measured gains sit in the code that plans its moves and the cameras that find the part.

Classical controllers have earned their place. When an arm's dynamics are fully modeled, computed-torque and model predictive control give mathematically predictable stability and precise trajectories. The review also notes their weak spots: joint elasticity, outside disturbances and thermal drift throw a rigid model off unless engineers retune it extensively by hand. Data-driven methods are where the authors find the quantified improvements.

Treat each figure as a ceiling. Each comes from a different task: trajectory planning, grasping in a crowded bin, calibrating geometry, recognizing scenes. They don't add up, and none of them predicts what a given line will see. They do point to the same place, and it isn't the gearbox.

The calibration result ties back to the serial-arm problem. Hybrid particle swarm and ant colony optimization cut geometric calibration error, which is the error a long six-axis arm accumulates. The review also reports that neural-network inverse kinematics solvers converge faster than standard Jacobian-based methods. That affects how quickly an arm can work out its joint angles for a new pose.

Perception is the other half. In structured cells, 2D and 3D cameras handle localization, dimensional inspection and pick-and-place alignment, and deep learning vision makes recognition more reliable in cluttered factory conditions. For micro-manufacturing and electronics packaging, the review describes 3D vision paired with LiDAR reaching millimeter-level localization.

The classical controller still has its place. The same review credits model-based control with predictable stability when the dynamics are known, and the fenced, deterministic jobs it describes are the cases where they're known best. A cell welding one part on one path all shift may get little from the reinforcement learning numbers. Those numbers matter to operators with mixed parts, cluttered bins or frequent changeovers, where rigid models struggle.

Qualcomm, Cognex and Omron bought the same layers

This fall's deals line up closely with the review's middle layers. Control Engineering's October 2026 roundup from investment bank Bundy Group counted 19 automation transactions, several of them dated from mid-September. Three of them land squarely on motion software, perception and compute.

Qualcomm agreed on Sept. 18 to acquire PickNik, longtime steward of MoveIt, the open-source manipulation framework built on the Robot Operating System. SME's Smart Manufacturing reported that MoveIt will stay open source under its existing license, with community-driven roadmaps and continued support for third-party hardware, while gaining tighter integration with Qualcomm's Dragonwing robotics platforms. PickNik's commercial MoveIt Pro is slated for the same integration. The stated aim is a smoother path from vision-language and vision-language-action models into planning, manipulation and real-time control. Closing is subject to customary conditions.

Cognex agreed on Sept. 22 to acquire RealSense, the depth-camera and 3D vision maker. The deal extends Cognex from industrial machine vision into robotic perception for fixed arms, mobile robots and humanoids. On Oct. 13, Omron agreed to acquire OnLogic and its parent, Interlogic, adding rugged industrial edge computers for compute-heavy manufacturing work.

Side by side, the three deals cover motion planning, 3D sensing and the computer that runs both next to the machine. That indicates suppliers expect cells to be judged on those layers. It also gives automation leads a practical question. If the planning framework, the depth camera and the edge box each come from a company with its own platform, openness between them belongs in the specification.

Lab results and the plant floor

The strongest objection to treating the review as a buying guide comes from two other recent reviews. A critical review of 82 studies on human-AI digital twins, published Oct. 3 in The International Journal of Advanced Manufacturing Technology, found that most work is still validated at laboratory scale. That finding is about digital twins, not manipulators, but it's a fair warning about any figure taken from the research literature.

A review in Advanced Science, published Oct. 26 by Kang Wang and colleagues at Zhejiang University and The Chinese University of Hong Kong, makes a related point. Much manufacturing AI still sits outside the physical system, and its robustness, interpretability and transferability in complex industrial settings are limited. Transferability is the issue that matters for a buyer. A grasping policy that works on one cluttered bin has to prove itself again on different parts in a different cell.

That gives engineers a straightforward test for any vendor's success rate. Ask which objects were used, how much clutter, how many trials and what counted as success. A 90% grasp rate from a clutter benchmark and a 90% rate from a month on a production line are different numbers that happen to share digits.

The next marker is the PickNik closing. Qualcomm has committed to deeper Dragonwing integration and to continued third-party hardware support. Automation buyers should watch how both commitments show up in MoveIt's release notes after the deal closes.