We as a team attended a number of robotics related talks at Nvidia GTC. We wanted to share some of our takeaways from the conference:
The Robotics Toolchain is Evolving Rapidly
While GTC only informs us about the tools of one company, that company happens to be the largest foundational force in AI/robotics.
What did we see?
ISAAC robotics platform took center stage as an end-to-end tool chain for robotics. The platform allows people to train, simulate, develop, and deploy software for their robots in an end-to-end fashion.
We learned about ISAAC ROS and its packages featuring hardware acceleration optimized for the Jetson and other Nvidia platforms. Isaac ROS also comes with the ability to use:
- GEMs - Which allows for easier deployment of GPU accelerated algorithms on ROS 2
- NITROs - Pipelines that allow for more efficient CPU/GPU utilization which greatly simplifies the process of building efficient heterogeneous computing onboard your robot.
Sidenote: We love to see the ROS 2 support from Nvidia!
Simulation is Eating Robotics the Same Way it Did For Semiconductors
Today, the semiconductor industry is blessed with the ability to do almost all of their development in simulation, which brought software-like speed to an industry that didn’t always have it.
Historically, in the early days of Fairchild Semiconductor, HP, and Intel, if you were building chips, the only way to do it was through physical trial and error: you would have to design your chip, fab it, and iterate over and over and over again. Each iteration was slow and expensive.
Remind you of any industry today? (hint:🤖)
Nvidia’s Isaac Sim seeks to change this. True photorealistic robotics simulations had us doing double-takes. Short of some lighting differences, it was nearly impossible to differentiate the photorealistic simulation from actual video footage. Synthetic data can be used to augment a company’s data sets, allowing development to kick off before sufficient real-world data can be collected.
Isaac Gym can help jumpstart reinforcement learning and get a robot trained considerably faster than other methods in some cases.
The Capital Needs to Start a Robotics Company are Following Moore’s Law
OK, so we didn’t graph this out exactly, but as compute gets cheaper and more accessible, it’s realistic to rapidly prototype in simulation more and more:
- The amount of physical prototyping required and, perhaps more importantly, the time between these revisions is decreasing, and cost to develop and deploy will follow. This is directly driven by the closing of the Sim2Real gap.
- The tooling acceleration will continue to push robotics hardware development toward the speed of software development.
- Synthetic data generation avoids the high operational costs of collecting enough data to train your models.
If you’re a robotics company in today’s world, we’d strongly suggest looking at Nvidia’s offerings. It’s in your best interest to see if they can accelerate your go-to-market or your existing robotics development process.
GTC Talks We 👍’d
Deep Reinforcement Learning with Real World Data
Real-World Implementations of Simulation for Next Generation Robotics
Isaac Sim: A Cloud-Enabled Simulation Toolbox for Robotics
Ready Robotics Accelerating their development with Isaac SIM (We ❤️ our friends at Ready Robotics)
