Nvidia recently unveiled Eureka, an AI system with the capability to autonomously train robots in executing novel tasks. During internal evaluations, Nvidia‘s Eureka successfully taught 10 virtual robots a total of 29 tasks, ranging from opening drawers to performing pen spinning tricks.
The method powering many robots is a neural network system known as reinforcement learning (RL). In this approach, robots engage in repetitive attempts within a virtual environment until they master the required task. Crucially, these AI training sessions are typically governed by reward functions. Nvidia’s Eureka stands out because it can automate the once arduous process of writing these reward functions, even transforming simple commands like “teach the robotic arm to play chess” into functional reward functions using OpenAI’s GPT-4.
Furthermore, Eureka’s system not only produces but continually refines these reward functions. It achieves this by developing several versions and assessing their efficiency through application on a simulated robot. The system also incorporates feedback from developers to enhance a robot’s reward function, leading to a notable 52% improvement in robot performance.
Nvidia has also made significant parts of Eureka accessible on GitHub, allowing developers to utilize the system alongside Nvidia’s Isaac Gym software.
IBM’s Energy-Efficient Breakthrough
IBM has taken the limelight with NorthPole, a unique AI chip boasting impressive energy efficiency. NorthPole, birthed in IBM’s Almaden lab, is a substantial advancement from its predecessor, TrueNorth, clocking in speeds approximately 4,000 times faster.
Key features of the NorthPole include its 22 billion transistors, streamlined into 256 cores capable of performing over 2,000 calculations every clock cycle. Unique to NorthPole is its architecture where the chip’s memory circuits are in close proximity to its cores, ensuring speedy data transfer and subsequent performance boost.
Dharmendra Modha, IBM’s chief scientist for brain-inspired computing, emphasized that the NorthPole chips blur the conventional distinction between compute and memory. Furthermore, the minimized data movement between the chip’s memory and cores not only ensures rapid processing but significantly conserves energy. In comparisons using the ResNet-50 model, NorthPole demonstrated 25 times better power efficiency than various graphics cards and CPUs.
Additionally, the processor’s enhanced energy efficiency translates to a cooler operational temperature. Thus, IBM’s NorthPole can operate with rudimentary cooling solutions, making it suitable for space-restricted settings like autonomous vehicles.
Bridging the Innovations
The leaps made by both Nvidia and IBM in the AI domain hint at a future where efficient robotic operations are merged seamlessly with enhanced computational powers. Nvidia’s progress in AI training harmonizes with IBM’s strides in power-efficient chip design. Both companies, in their unique ways, are setting the stage for a future where robots are not only smarter but also energy-conscious. It’s not just about doing more; it’s about doing more efficiently.