New Era of Robotics: Intel’s Song Jiqiang on Embodied Intelligent Robots and Their Computational Needs

In the realm of artificial intelligence, computational power, algorithms, and data are the three core elements. These aspects are even more crucial for embodied intelligent robots. Among these elements, the development of computational infrastructure has progressed the furthest, encompassing various aspects such as technology, processes, and industrial ecology. However, there is a notable scarcity of chips specifically designed for embodied robots. This is largely due to the current immaturity of the embodied robotics industry, which lacks unified standards and the capacity for large-scale production. As a result, many solutions are still handcrafted, which is a common situation in this sector.

In this context, the second episode of the New Era of Robotics podcast featured a special guest, Song Jiqiang, Director of the Intel China Research Institute. Joining him was Gao Fei, CEO and Editor-in-Chief of Zhidian Technology, to discuss the computational requirements for embodied intelligent robots. Below are some selected Q&A highlights from the episode.

Q&A Highlights

1. Early Humanoid Robots Required Large Support Teams

Q: You previously worked at the Robot Interaction Research Center. Can you explain what this institution does and its main focus of research?

Song Jiqiang: The center was established around 2014, coinciding with a global maker movement and a surge in maker activities in China. At that time, intelligent robots were still in their infancy, and the focus was on developing their sensory and voice interaction capabilities. It was crucial for robots to understand human language and respond appropriately, either through actions or movements. Intel had released several compact computing modules combining computation, communication, and storage, which were utilized by various robot and drone manufacturers. By 2014, we had established a Robot Interaction Innovation Research Center to enhance the interaction experience between humanoid, semi-humanoid, and even pet-like robots and humans.

Q: I recall that this period coincided with the rise of smart voice technologies, particularly NLP. Was the research center primarily focused on voice recognition technology at that time?

Song Jiqiang: Voice recognition was certainly a crucial focus, alongside visual recognition. Between 2012 and 2014, deep neural networks made significant strides due to ImageNet, allowing cameras to capture various environmental scenes and enable interaction with humans. We integrated visual and voice recognition as a cohesive focus in 2014.

2. AI and Multi-Sensors: Essential for Humanoid Robot Capabilities

Q: In 2013, Intel established a perceptual computing product line, leading to what we now know as RealSense. What drove that decision?

Song Jiqiang: The emergence of new AI technologies enhanced visual input understanding and detection capabilities, prompting us to explore whether this could boost robots’ capabilities. The robot industry has a long development history, predating both computing and AI. Every new technology raised the question of how it could be applied to robotics, especially humanoid robots. Thus, we continuously integrated new technologies and conducted experiments. The advancements in visual neural networks around 2012 and 2013 necessitated that sensors keep pace; robots needed to comprehend their environment in three dimensions to interact effectively.

3. Learning in Robotics: From Programming to Deep Learning

Q: How have deep learning and imitation learning changed research paradigms in robotics?

Song Jiqiang: Initially, control models like MPC were pre-designed, making learning a human task. With deep learning, we began to explore how robots could learn actions that were previously challenging to program directly. Imitation learning and reinforcement learning became significant, enabling robots to observe and replicate predefined action sequences either through video or simulations. This shift allows robots to converge more quickly towards usable action paths, although not necessarily optimal ones.

4. The Future of Embodied Robots and Their Computational Needs

Q: What level of computational power do you think embodied robots will need, especially in industrial applications?

Song Jiqiang: The computational needs vary across different industrial sectors. For basic tasks like picking and placing items, the required power is mainly for visual processing and VLA models, generally under 200 TOPS for optimized models. However, custom models may demand more power. We can enhance capabilities with external AI acceleration cards or low-latency wireless communication linked to edge computing boxes, especially when multiple robots operate in a shared environment.

This podcast episode provides valuable insights into the evolving landscape of embodied intelligent robotics and the computational needs that will drive this technology forward.