Large AI Robotics Simulator for Robotics Applications

Challenge

Developing robotics applications involves software, simulation, control systems, machine learning, physical testing, and a large amount of uncertainty. There’s a need for a platform that can bring together these disparate systems and combine them to provide an environment for meaningful testing, validation and verification.

Robotic applications are subjected to inherently risky situations. The environments that they are deployed in are often harsh and remote, making testing challenging. Their complex nature also requires several different systems to be tested and validated to work together, often in novel ways. What if there was a way to test, validate and verify robotics virtually?

Solution

LARS Sim (Large AI robotics simulator) brings several technologies into a single platform for robotic development.

The platform includes:

• High-fidelity graphically accurate environments that represent what is ‘seen’ by robotics systems
• Ability to test Robotics Operating System (ROS) in the platform
• Annotated data generation for machine learning applications
• Ability to run huge scale simulations in parallel using HPC
• Ability to validate simulations through physical robotics.

Impact

LARS Sim enables robotic behaviour to be explored in a representative virtual environment before moving onto physical testing. This enables the evaluation of performance, the investigation of different operating conditions and the refinement of software.

Through running parallel simulations on high-performance compute infrastructure, we can enable thousands of simulations to be run, creating large amounts of high-fidelity data to help train machine learning models to support autonomous applications such as defect detection and foreign object detection. We’ve also created the platform to ensure that robotics simulation
can be run on any form of GPU, providing greater flexibility when selecting compute hardware. It also enables users to use AI to help generate large number of test scenarios in an optimal timeframe, allowing for better testing