Physical AI Moves from Research into Industrial Automation
Agile Robots is pushing Physical AI from laboratory research toward practical industrial automation.
In August 2026, the company demonstrated this direction at events in Switzerland and Germany.
Its demonstrations connected intelligent robots, sensing technology, automation software, and robot training data.
Therefore, the company’s strategy extends beyond conventional robotic motion and factory automation.
Intelligent Robotics at All About Automation Zurich
At All About Automation Zurich, Agile Robots presented solutions for manufacturers and system integrators.
The demonstrations focused on assembly and welding applications with practical production requirements.
Moreover, the company showed how sensing and software can simplify robotic deployment.
Diana 7 Applies Force Control to Assembly
The Diana 7 robot uses torque sensing across all seven joints for force-controlled assembly.
During the demonstration, the robot inserted an engine head while monitoring mechanical forces continuously.
This approach helps detect insertion conditions and reduces the likelihood of jamming or component damage.
For industrial automation, force feedback can complement conventional position-based robot programming.
From an engineering perspective, this capability matters when components require controlled contact during assembly.
Traditional automation often depends on precise fixtures, tolerances, and carefully defined motion paths.
Force sensing can provide additional process information when mechanical variation affects the assembly operation.
Thor 12 Targets Flexible Robotic Welding
Agile Robots also demonstrated the Thor 12 robotic welding platform at the Zurich event.
The system supports drag-and-drop teaching and low-code programming for welding applications.
Operators can create welding paths and modify them without developing extensive conventional robot code.
In addition, the robot can address corner, vertical, and inclined welding positions.
The demonstration included welding conditions with gaps as narrow as 1 mm.
Such flexibility can benefit manufacturers that frequently change products or welding geometries.
However, production results still depend on material preparation, torch setup, process parameters, and welding quality controls.
Robot Training Data Becomes a Core Physical AI Requirement
Franka Robotics demonstrated another part of the Physical AI ecosystem at IJCAI-ECAI 2026.
The event took place in Bremen from August 15 through August 21, 2026.
As a Silver Sponsor, Franka Robotics presented technology for collecting robot-learning demonstrations.
Visitors used the Franka GELLO Duo to teleoperate the FR3 Duo bimanual robot.
The LABS platform captured these demonstrations as structured training data for manipulation tasks.
This workflow connects human operation, robotic hardware, and data collection within one environment.
Bimanual Robotics Supports Robot Learning
Bimanual manipulation presents different challenges from conventional single-arm robotic automation.
A robot must coordinate two arms while maintaining timing, position, force, and object relationships.
Consequently, high-quality demonstration data becomes important for developing capable robot-learning models.
Human teleoperation provides a practical method for generating such demonstrations.
Instead of manually programming every movement, operators demonstrate desired behaviors through robotic interfaces.
The resulting datasets can then support research into learned manipulation and embodied AI.
Physical AI Extends Beyond Traditional PLC and DCS Architectures
Physical AI does not replace PLC, DCS, or established control systems in every application.
Instead, it can complement deterministic automation where perception, adaptation, and complex manipulation create additional requirements.
PLCs remain well suited to deterministic sequencing, interlocks, machine control, and safety-related functions.
DCS platforms continue to manage process control, instrumentation, alarms, and plant-wide operational functions.
Physical AI can add another computational layer for perception and adaptive robotic tasks.
Therefore, future factory automation may combine conventional control architectures with intelligent robotic systems.
This hybrid approach can preserve deterministic control while adding flexible machine-learning capabilities.
Industry Perspective: Data Quality May Define the Next Robotics Advantage
From an industrial automation perspective, robot hardware alone will not determine future Physical AI performance.
Training data quality, process consistency, sensing capability, and deployment infrastructure will also influence results.
Manufacturers should therefore evaluate the complete automation workflow rather than focusing only on robot specifications.
They should examine data collection, integration, programming, validation, maintenance, and operator interaction.
Moreover, production teams must define measurable process targets before deploying learning-based robotic solutions.
This consideration becomes particularly important for factories with multiple products and frequent production changes.
Flexible robotic systems can reduce engineering effort when product variation remains manageable.
However, each application still requires appropriate validation, process controls, and production acceptance criteria.
Application Scenario: Adaptive Assembly in Automotive Manufacturing
An automotive supplier could use force-controlled robotics for component insertion and alignment.
The robot could monitor joint torque while approaching the assembly position.
If the measured force differs from expected conditions, the system could adjust its motion or stop.
This strategy can reduce mechanical stress and provide additional process information.
A PLC could continue managing machine sequencing, permissives, sensors, and safety-related interlocks.
The robot controller would then handle motion and force-control functions within its defined operating envelope.
This architecture demonstrates how Physical AI can complement existing industrial control systems.
Application Scenario: Flexible Welding for Mixed Production
A welding manufacturer could deploy a flexible robot for short production runs and changing product geometries.
Low-code programming can reduce the effort required to create new welding paths.
Operators can therefore spend less time rebuilding conventional programs for every production change.
Nevertheless, welding quality requires more than robot path flexibility.
Manufacturers must control joint preparation, wire parameters, shielding gas, torch angle, and thermal conditions.
Therefore, robotic intelligence should operate alongside established welding process controls.
From Robot Hardware to Complete Automation Platforms
Agile Robots and Franka Robotics illustrate a broader change within industrial robotics.
Robot manufacturers increasingly combine mechanical platforms, sensing, software, data, and artificial intelligence.
This development moves factory automation toward systems that can perceive and adapt to physical environments.
For automation engineers, the key question is not whether Physical AI replaces PLC or DCS technology.
Instead, the more practical question concerns where intelligent robotics creates measurable production value.
Applications involving variable components, complex manipulation, frequent changeovers, and limited programming resources may benefit most.
Conclusion: Physical AI Needs a Practical Industrial Path
The demonstrations in Zurich and Bremen show two complementary directions for robotic technology.
Diana 7 and Thor 12 address practical industrial automation requirements such as assembly and welding.
Franka Robotics focuses on collecting demonstrations that support future bimanual robot-learning applications.
Together, these developments highlight the growing connection between robotics, artificial intelligence, and industrial control.
The next stage of factory automation will likely combine deterministic control with adaptive robotic intelligence.
For manufacturers, successful deployment will depend on measurable benefits, suitable integration, and disciplined engineering validation.