The Silver Age of Embodied Intelligence: Exploring Robotics with Large Models

The Silver Age of Embodied Intelligence: A Deep Dive

The field of embodied intelligence is currently experiencing what is being called a "silver age," a period characterized by intense exploration and development, situated between the initial emergence and the full maturation of the technology. This phase is marked by significant advancements and a growing understanding of how to integrate AI with physical systems to create more intelligent robots. The discussions surrounding this exciting period often center on the potential of large AI models to accelerate the development of robotics and the challenges that need to be overcome to fully realize this potential.

Background: The Robotics Boom and its Challenges

The rapid progress in large AI models has spurred massive investment in the robotics industry. This surge in funding, while indicative of the field's promise, also introduces the risk of market overheating. The central challenge lies in distinguishing genuine technological breakthroughs from mere hype and identifying viable applications that can make a significant impact. This involves addressing fundamental questions about the best paths forward:

Should the primary focus be on reinforcement learning, or should simulation learning take precedence?
Is it more crucial to prioritize simulation environments or real-world testing when developing robotic systems?
Should the emphasis be placed on vision systems or physical engines in the design of advanced robots?

These questions highlight the complexities inherent in the development of embodied intelligence and the need for a strategic approach to navigate the current landscape.

Roundtable Insights: Experts Weigh In

A roundtable discussion at the Volcano Engine FORCE conference provided valuable insights into these questions. Experts from diverse fields gathered to explore the potential of large models in accelerating the development of robotics. The participants included:

Chen Yang: Vice President of Galaxy General Robotics
Shi Lingxiang: Head of Innovation Incubation at Volcano Engine (Moderator)
Wu Di: Head of Intelligent Algorithms at Volcano Engine
Wan Haoji: Partner at Matrix Partners China
Wang Xiao: Founder of Nine Chapters Capital
Yan Weixin: Co-founder of Shanghai Zhiyuan Robotics and Doctoral Supervisor at Shanghai Jiao Tong University

Their collective expertise offered a comprehensive view of the current state of robotics and the potential pathways for future growth.

The Surge in Robotics Investment: Why the Hype?

The excitement surrounding robotics stems from its unique position within the broader field of AI applications. AI applications can be broadly divided into two categories: soft applications, such as chatbots and video generation, and hard applications, which include robotics. Robotics is considered the most versatile hard application of AI, as it has the potential to impact a wide range of industries and applications.

Investors are particularly interested in companies that can effectively integrate both software and hardware and demonstrate real-world applications that go beyond simple demos. The focus is on tangible results and the ability to deploy robots in practical, beneficial ways. However, the commercialization of robots has been slower than expected, especially in complex environments like homes and B2B services.

Several factors contribute to these challenges:

The need for better coordination between the "brain" (AI) and the "small brain" (control systems) of robots.
The critical importance of cost reduction to enable widespread adoption.

These factors highlight the need for both technological advancements and strategic business models to push the field of robotics forward.

The Path to Commercialization: Navigating Uncertainty

While there's a strong consensus that robotics will eventually be successful, the exact timeline and the companies that will lead the way remain uncertain. It's unlikely that a single company will dominate the market, similar to the electric vehicle industry, where multiple players have found success.

Large models have already improved robots' interaction and thinking capabilities. However, there are still significant technical hurdles to overcome. These hurdles, while not insurmountable, are expected to make the process longer and more challenging than initially anticipated. Venture capitalists play a vital role in accelerating this development by providing the necessary funding for research and development.

The Importance of General Intelligence: Shifting Perspectives

A key shift in focus is the emphasis on robots adapting to humans and environments rather than the other way around. This requires robots to possess a higher degree of general intelligence, which can be achieved through:

The use of large amounts of simulation data to train robots.

However, robotics startups face significant challenges in technology, product development, and business model creation. These challenges underscore the need for:

Increased collaboration across the supply chain
Stronger support from investors

Technical Paths for Embodied Intelligence: Exploring Different Approaches

Several technical paths are being explored to advance embodied intelligence. These include:

Imitation and Reinforcement Learning: Combining imitation learning to enhance reinforcement learning is proving to be a viable approach for gait control.
Simulation for Lower Limbs: Simulation data has shown to be effective for lower limb gait control, although parameter tuning and product consistency remain challenges.
Focus on Upper Limbs: There is a growing recognition of the need to shift focus from lower limb movement to the overall task operation capabilities of humanoid robots.
Task Operation: The emphasis should be on a robot's ability to perform complex tasks rather than just locomotion.

Data is a critical component in the development of embodied intelligence. The field faces several data challenges:

The difficulty of collecting and standardizing data, particularly for complex tasks.
The crucial role of real-world data, especially for complex physical interactions that are difficult to simulate.

Simulation vs. Real-World Data: A Balancing Act

The debate between using simulation data and real-world data is ongoing, with each having its own advantages:

Simulation Data: It's more cost-effective, scalable, and versatile for training general-purpose embodied models.
Real-World Data: It's essential for capturing the nuances of physical interactions, such as friction and elasticity, which are hard to replicate in simulations.

The development of reliable world models is seen as a key enabler for large-scale simulations to be used effectively to test and improve robot performance in a variety of scenarios.

Future Applications: Near and Long-Term Visions

The future of embodied intelligence is filled with possibilities, with applications spanning various sectors:

Near-Term Applications (2-3 Years)

Industrial Manufacturing: Robots can perform complex tasks requiring dexterity in controlled environments.
Remote Operations: Robots can be used in hazardous environments, such as handling dangerous materials.
Controlled Environments: Robots will be deployed in controlled environments like restaurants, hotels, and factories.
Specific Tasks: Robots will be used for tasks like delivering food, making coffee, and performing light maintenance.
Initial Deployment Areas: Factories, offices, and security are the most likely areas for initial deployment.

Long-Term Applications

Home Environments: The most complex yet highly anticipated application is in home environments.
Household Tasks: Robots will eventually be able to perform tasks like cooking, folding laundry, and cleaning.
Cost Reduction: As technology advances, the cost of robots will decrease, making them more accessible to consumers.
General-Purpose Robots: The focus will shift towards general-purpose robots that can serve a variety of needs.
Market Considerations: Companies need to consider the functionality, performance, openness, and risk tolerance of different applications.

Volcano Engine VeOmniverse: A Platform for Innovation

The Volcano Engine VeOmniverse is a virtual simulation platform designed to accelerate the development of robotics. It provides a cloud-based environment for robot simulation and training.

Key features of the platform include:

Realistic Environments: It creates highly realistic digital environments for training and testing robots.
Cost-Effective: It reduces the need for physical equipment, thereby lowering development costs.
Comprehensive Training: The platform integrates visual engines, physical engines, sensor simulation, and 3D generation to create a comprehensive training system.
AI Support: The platform uses AI to generate high-quality training data and accelerate the training process.
Customization: The platform is open and customizable, allowing companies to develop personalized digital twin applications.
Accelerated Development: It helps companies quickly build, validate, and optimize robot models.
Industry Transformation: VeOmniverse is a key tool for the intelligent and digital transformation of the robotics industry.

This platform underscores the importance of simulation tools in the ongoing development of embodied intelligence and the transformative potential of these technologies.