Electronics Era: With the rise of embedded AI and smart machines, how does AuraML enable engineers to simulate and validate physical AI systems at the design stage itself?
Ayush Sharma: AuraML’s multi modal world simulation model generates hyper realistic, fully physics ready 3D scenes using text prompts, floorplans or real world video or sensor data. Accurate sensor simulation for Lidar, IMU and cameras helps bridge the sim to real gap and the automated agentic orchestration helps engineers test and train Physical AI systems without much manual effort. Without the need of any real world hardware they can understand how their robots will perform right from the design phase all the way to final real world deployment.
Electronics Era: Traditional testing of robotics and industrial systems is hardware-intensive. How does simulation-first development reduce iteration cycles and hardware dependency?
Ayush Sharma: Traditional testing requires months of real world tests in live locations. The robotics companies need to physically build replicas of customer facilities or get special permission to test on live customer locations. The engineers to travel on site, robots and other hardware needs to be transported as well. Every testing scenario needs to be manually set up and executed.
AuraSIM first solves this bottleneck as thousands of variations and iterations can be tested in a few days. Every single edge case and failure scenario can be tested, which is impossible to do in the real world. Our world model can generate all possible scenarios and do large scale multi robot, multi agent tests in a few hours by just typing a few prompts instead of the complex and time consuming real world tests.
Electronics Era: How does AuraML’s platform help engineers model real-world variability—such as sensor noise, environmental conditions, and system failures—within simulation environments?
Ayush Sharma: AuraSIM has its own proprietary sensor simulation models that can accurately model the customers’ sensors. By fine tuning on just a few hours of real sensor data the model can learn to output data that matches the real world sensors exactly.
The multimodal world simulation model can generate 3D scenes that are accurate and generate failure cases, edge cases, varying environmental conditions with just a prompt. The user can just say: “Make the boxes in the rack fall when the robot passes in front of it” and the model will create a simulated scenario with a described edge case.
Electronics Era: Edge AI is becoming critical in industrial and electronics systems. How does AuraML support training and validation of AI models that are eventually deployed on edge devices?
Ayush Sharma: Our multi modal world model has been built from the ground up to function as a teacher model for these Edge AI use cases. The smaller models that will be deployed on the edge learn from the simulated scenarios generated by the world model, our integrated RL training pipelines generate accurate simulated data for the edge to learn from. AuraSIM offers parallel training where the users can create thousands of copies of the robot in the cloud and train them parallely converging on a stable RL policy which can be deployed on the edge.
Electronics Era: From an engineering standpoint, how are machine learning models trained within simulated environments to replicate real-world physical behavior and system dynamics?
Ayush Sharma: The simulated environments generated by the multi modal world model behave exactly similar to the real world. The physics properties and the object interactions are modeled perfectly. The sensor simulation model accurately outputs simulated sensor data that is perfectly similar to the real sensor. The agentic cloud orchestration can generate hundreds of simulated worlds with a lot of variations which helps train the RL policies.
The users simply import their robot 3D/CAD model into AuraSIM which automatically creates a sim ready version of the robot. The engineers then define a policy with a reward function that rewards correct behaviour. Finally they set up simulation runs with the exact scenarios they want their robot to function in and train it on hundreds of parallel simulation runs. This allows the machine learning model to do years worth of training in a day or two with the rich variations and dynamics expected to see in the real world.
Electronics Era: How does AuraML handle large-scale, high-fidelity simulation workloads while ensuring performance and scalability for industrial use cases?
Ayush Sharma: The power of modern scalable Nvidia GPU clusters and the Omniverse framework allow us to handle large scale workloads.
Ayush Sharma: Our simulation framework can utilize hundreds of GPUs in parallel to split the large workloads and ensure performance and scalability. Our simulation agent can effectively orchestrate and manage these high fidelity simulations on multi gpu clusters.
Electronics Era: One of the key challenges in AI systems is robustness. How does simulation help identify failure modes, edge cases, and system vulnerabilities before deployment?
Ayush Sharma: AuraSIM has features that allow users to easily create rich 3D scenarios using different inputs. We also have randomisers and variation generators built in. The users simply run their physical AI systems in all possible scenarios.
In the real world we are limited by the things we can do, in simulation we can make anything happen, make accidents happen, fires, safety hazards etc. without putting any real hardware or people at risk.
Electronics Era: How does AuraML integrate with existing engineering stacks such as robotics frameworks, embedded systems, or industrial automation workflows?
Ayush Sharma: AuraSIM integrates seamlessly with existing frameworks like ROS, VDA5050 and others. We also integrate directly with the users CI/CD workflows using Github or Jenkins. We also offer importing data from various 3D formats and CAD tools etc into our world simulation model directly.
Electronics Era: What are the most common gaps you see today between AI model development and real-world deployment in physical systems, and how does AuraML bridge that gap?
Ayush Sharma: The most common gaps we see in the physical AI models today are:
- Limited generalisation: Models are very specific to the environment they are trained in any slight change in lighting, textures or objects causes the model to fail
- High failure rates: Models are trained on limited data have high failure rates and need constant manual intervention
- Need re training: For every new customer or even different facilities of the same customer require the process to be repeated, new data and retraining
- No context/integration: The models are isolated boxes which cannot learn from surrounding things or integrate with the warehouse management systems.
- No multi agent: Current models are single entities that work alone, but rarely there is one robot but many working together in sync
Electronics Era: How do distributed simulation and compute architectures improve the efficiency of training physical AI systems at scale?
Ayush Sharma: Training physical AI systems at scale requires massive GPU clusters and the ability to run 1000s of parallel training runs.
The more parallely the architecture can scale, the faster the models can be trained.
Electronics Era: As industries move toward autonomous systems and smart manufacturing, what role will simulation platforms play in the future of electronics and system design?
Ayush Sharma: In the future every aspect of smart manufacturing will be validated first in simulation before being built in the real world.
Right from designing the factory floor to deploying fleets of autonomous robots everything will be simulated first. Only once things are confirmed to be working in simulation will the real world construction or deployments progress. This will result in massive efficiency improvements all around.
Electronics Era: Traditional simulation tools have been used in engineering for years. How does AuraML differentiate itself from conventional simulation platforms, particularly in the context of training AI-native, real-world physical systems?
Ayush Sharma: Traditional simulation tools require a lot of manual effort to make them work and are often limited in what they can do and have a large gap to reality.
They need a large in-house team to maintain and run and need a lot of interdisciplinary expertise.
AuraSIM has the following advantages:
- Multimodal world model: Generate physics and sim ready 3D scenes from prompts, floorplans or videos. Traditionally this is done by humans by hand which takes months for a single scene. Our model can generate hundreds of them in a day.
- Sim to real gap: Our accurate sensor simulation and GPU powered physics engine have a very low sim to real gap.
- Agentic orchestration: Our sim agent can generate and run the simulation for you. Traditionally the simulations require a large team of engineers to maintain and run
- Automated integration: Integrateds with the users pipeline directly and enables them to run hundreds of simulations parallely with ease.
