Industry
Manufacturing
Future-Ready Manufacturing: Partnering in your journey to a sustainable manufacturing enterprise
Highlights
- Industrial robotics is moving beyond fixed, pre-programmed motions toward systems that can perceive, learn and adapt to variations in the physical environment.
- TCS and Volvo are exploring a reinforcement-learning-based approach for adaptive pick-and-place using Universal Robots UR10e, combining perception, motion planning and learned robotic policies.
- NVIDIA Omniverse libraries and the open Isaac Sim and Isaac Lab learning and simulation frameworks provide a physics-based digital environment to train and validate robotic policies across variations in object position, orientation and type before physical deployment.
- The approach creates a pathway toward faster robot training, reduced physical experimentation and more adaptable automation, with performance measured across simulation, policy convergence and eventual sim-to-real execution.
On this page
On this pageinpage
From programmed motion to adaptive robotic intelligence
Industrial robots have transformed manufacturing by bringing precision, repeatability and speed to production environments. Yet many robotic applications remain highly dependent on structured work cells and carefully programmed trajectories. When the position, orientation or configuration of an object changes, engineers may need to modify motion programs, tune parameters or introduce additional logic.
As manufacturers look to deploy robots across more dynamic environments, this model presents a new challenge: the robot needs to respond to variation rather than simply execute a predefined sequence.
This is where Physical AI is beginning to change the robotics equation.
Instead of explicitly programming every possible motion, AI-based approaches can allow robots to learn how to respond to different states of their environment. For manipulation tasks, reinforcement learning is particularly relevant because the robot must continuously translate information about its state and the objects around it into physical actions.
TCS and Volvo are exploring this approach through an adaptive pick-and-place capability for a six-degree-of-freedom Universal Robots UR10e. The solution combines visual perception, reinforcement learning and motion planning, with NVIDIA Omniverse libraries providing the simulation environment in which robotic behavior can be developed and evaluated.
Why reinforcement learning for robotic manipulation
A conventional pick-and-place application can be straightforward when the object arrives at a known location and orientation. The complexity increases significantly when those conditions change.
An object can appear at different points within the robot’s workspace. Its orientation can vary. The robot’s approach trajectory may need to change accordingly. A solution that relies entirely on predefined trajectories must account for these combinations explicitly.
Reinforcement learning provides another approach. Instead of programming each individual response, a policy can learn a relationship between the observed state and the action required to accomplish the task.
For the UR10e use case, the learned policy is developed to support adaptive manipulation while operating within defined constraints. The approach does not replace conventional robotics algorithms. Instead, reinforcement learning works alongside perception and motion planning, with each component addressing a different part of the problem.
Perception determines what is where. The learned policy determines how the robot should respond. Motion planning determines how to execute the movement safely and feasible. This combination creates an architecture better suited to environments where the robot needs to respond to variation rather than follow a single fixed trajectory.
Teaching robots in a digital world
Training reinforcement-learning policies directly on physical robots can require substantial time, hardware availability and repeated experimentation. It can also introduce safety and equipment constraints when a policy is still learning.
TCS and Volvo are addressing this challenge through a simulation-first development approach using NVIDIA Isaac Sim and Isaac Lab for simulating and training.
A digital twin of the UR10e work cell represents the robot, two-finger gripper, camera, pick and drop bins, objects and surrounding environment. Physics-based simulation allows the robot to interact with these elements while the training environment introduces controlled variations.
Object locations can be randomised across the workspace. Object orientations can be varied. Different object types can be introduced into the environment. These variations expose the learned policy to a broader range of conditions during training and validation.
Domain randomisation is particularly important for Physical AI because a policy trained against a single idealised configuration can become overly dependent on that environment. Introducing controlled variation during simulation encourages the policy to learn behavior that is more robust to changes.
The simulation therefore becomes more than a visualisation of the robot. It becomes a development, training and validation environment for robotic intelligence.
From perception to physical action
An adaptive robot also needs to understand its physical environment.
The solution integrates an Intel RealSense RGB-D camera for object detection and six-degree-of-freedom pose estimation. The depth information provides spatial context that can be used to determine where an object is located and how it is oriented.
The resulting perception information feeds into the robotic control pipeline. Reinforcement learning provides adaptive behavior, while NVIDIA cuMotion supports motion planning and trajectory generation.
The resulting sequence is consistent across the simulated and physical environments:
Perceive → estimate pose → pick → move adaptively → place → return
This architecture allows the different capabilities to work together rather than treating AI as a replacement for the existing robotics stack.
The approach also creates a clear separation between learning and execution. The policy can be trained and evaluated extensively in simulation, while established motion-planning and robot-control mechanisms continue to provide the execution framework.
Measuring robotic intelligence before deployment
For Physical AI to move from experimentation to production, demonstrating that a robot can complete one successful task is not enough. The behavior needs to be measured across representative conditions.
The UR10e capability is therefore being evaluated across variations in object type, location and orientation, including more than five locations within the workspace and randomized object orientations.
The evaluation framework can capture metrics across several stages of the development lifecycle:
- Pick accuracy and task success: How consistently the robot detects, grasps and places the intended object.
- Policy convergence: How the reinforcement-learning policy improves during training and when performance stabilises.
- Simulation robustness: How performance changes as object position, orientation and other environmental variables are varied.
- Sim-to-real performance: How much of the validated simulation performance is retained when the learned behavior is transferred to the physical UR10e.
- Inference latency: How quickly the trained policy can generate actions during execution.
These measurements provide a more meaningful view of robotic intelligence than a single successful demonstration. They establish a quantitative basis for determining whether a policy is sufficiently robust to progress from simulation toward physical deployment.
Closing the simulation-to-real gap
The ultimate test for a simulation-trained policy is the physical robot.
The next stage of the work involves transferring the integrated perception and control pipeline to the Volvo UR10e physical work cell, including the physical robot, Intel RealSense camera, pick and drop bins and representative objects.
This transition requires more than transferring a policy file. The physical environment introduces real sensor characteristics, robot dynamics, hardware interfaces and operating constraints that must be accounted for during validation.
The deployment architecture therefore includes ROS2 integration, robot and controller interfaces, camera access and defined safety procedures. Real-world testing can then be performed against agreed scenarios and acceptance criteria.
The objective is to determine how effectively intelligence developed in simulation transfers to physical execution and where additional calibration or adaptation may be required.
This creates a measurable feedback loop:
Simulate → Train → Validate → Transfer → Measure → Improve
Such a loop can reduce the dependence on trial-and-error programming directly on physical equipment while providing a structured path for bringing AI-based robotic capabilities into industrial environments.
From one robotic task to a reusable Physical AI foundation
The significance of the approach extends beyond a single pick-and-place application.
The combination of digital twins, physics-based simulation, domain randomisation, perception, reinforcement learning and motion planning provides a potential foundation for developing other adaptive robotic capabilities.
The same development pattern could be applied to industrial scenarios where robots need to respond to variation in their environment, including material handling, kitting, machine tending, inspection and other manipulation tasks.
Business opportunities are equally important. By moving more training and experimentation into simulation, manufacturers can potentially reduce physical testing cycles, accelerate development and evaluate a broader range of operating conditions before deploying changes to production equipment.
The longer-term shift is from programming robots for every expected situation to developing robots that can learn to operate across a defined range of situations.
Working with Volvo, TCS is exploring this shift through simulation-first Physical AI architecture powered by NVIDIA Omniverse libraries. The work brings together AI-based learning and established robotics technologies to create a pathway from virtual training to physical execution — providing a foundation for more adaptive, measurable and scalable industrial automation.