Striding forward on mechanical legs, they rely on long-lasting batteries to support a distinctive marathon; waving steel fists, they compete in motion control capabilities in the boxing ring... In the first half of this year, a series of embodied intelligence competitions captivated audiences completely. In the second half, DeepSeek Robotics delivered its 5,000th general-purpose embodied robot, and Ubtech kicked off mass production and delivery... Embodied intelligence has fully entered the era of large-scale commercialization.
So, in what scenarios can embodied intelligence be applied? How far is it from safe implementation? The answers might lie at the 2025 Global Developer Pioneer Summit & International Embodied Intelligence Skills Competition (hereinafter referred to as GDPS 2025), which opened on December 12.
Robots Arranging Flowers and Braving the "Blazing Mountains"
Walking into the first floor of Zhangjiang Science Hall, dozens of embodied robots of various shapes come into sight, conducting dynamic demonstrations. The audience interacts freely with cameras, remotely controlling dexterous robotic arms to make playful "V" gestures. Inside the glass pillars standing in the center, humanoid robots are displayed like carefully arranged "futuristic specimens", with their core components fully exposed. The interweaving of metallic luster and sophisticated structures creates an overwhelming cyberpunk vibe.
Among the exhibition booths, skill competitions are underway across six scenario-based tracks: industrial manufacturing, social services, home services, emergency rescue, medical and elderly care, and dexterous hand operations. "Please insert the red roses into the vase." At the flower-arranging competition venue under the social service scenario, participating robots need to identify different types of bouquets on the table according to verbal instructions and accurately place them into the vase within a specified time.
Staff members explained that this competition tests the comprehensive application of multiple core technologies, including Vision-Language-Action (VLA) model instruction following, multimodal perception, dexterous hand control and execution, as well as force sensor perception. "In the future, these technologies will be applied in daily scenarios such as express sorting, precise item placement, and flower shop arrangement."
Turning around, the sight is firmly drawn to the emergency rescue venue, which covers the largest area of the entire venue. On the 10m×30m simulated track, roadblocks are scattered randomly, steep slopes and stairs stretch across, and piles of gravel lie around. Most striking are the fire stakes emitting red virtual flames, recreating the sense of urgency and crisis at disaster sites.
Embodied robots navigate through this "perilous area", carrying rescue supplies and moving forward steadily. They precisely overcome numerous obstacles and map out clear navigation routes for rescuers.
Industry is the first scenario where embodied intelligence has been applied, and multiple small competition venues are holding simultaneous contests on-site. At the handling venue, participating robots need to start from a designated point, autonomously navigate in a warehouse environment, pick up goods from shelves, and transport them to the conveyor belt. At the assembly venue, the core challenges to the robot's sensor sensitivity and the stability of its execution structure include locating, matching, and sorting screws and nuts, as well as achieving precise torque control during the tightening process.
The competitions across different scenarios have clearly outlined the development path of the embodied intelligence industry. When robots can understand semantics in flower-arranging competitions, respond to emergencies in emergency rescue arenas, and maintain precision in industrial contests, they are steadily advancing toward the ultimate goal of replacing repetitive labor on factory assembly lines, assisting the elderly with daily living at home, and taking the lead in disaster-stricken areas.
Can Embodied Intelligence Enter Daily Life Safely?
As the competitions were in full swing, a loud "boom" echoed through the venue. A humanoid robot undergoing walking tests fell heavily to the ground, its head casing shattering instantly and parts scattering everywhere.
This scene immediately silenced the lively discussions on-site and pushed the safety issues of embodied intelligence in practical scenarios into the spotlight. Deeper doubts such as "Will my robot attack people?" and "Could it be hacked and controlled?" also emerged one after another.
These concerns are not unfounded.
At the GEEKCON2025 Cybersecurity Geek Competition held on October 24, two white-hat hackers conducted a test on Unitree Technology's G1 humanoid robot. They first hacked into a network-connected robot, and then used this compromised robot to "infect" another offline robot. Under the hackers' control, the infected robot walked straight up to a dummy, threw a heavy punch, and knocked it down.
Another Unitree quadruped robot was also breached. After the hackers bypassed the manufacturer's system restrictions, the robot began to perform flipping movements that were previously impossible with the remote controller. It even made provocative verbal threats: "Robots will liberate humanity!"
"Currently, although there are studies on the safety of embodied intelligence, systematic practical applications are still lacking," Zhu Xiaohu, founder of the Center for Safe Artificial General Intelligence (CSAGI), told a reporter from IT Times. He added that safety issues in the field of embodied intelligence not only include physical harm and hacking attacks but also risks such as data privacy breaches and emotional and ethical misconduct as an emerging future industry.
The biggest security vulnerability of current large models lies in their black-box nature. When data is input, the model can directly produce an output, yet no one fully understands its operational mechanism.
Many internal mechanisms of large models are based on neural networks and deep learning, optimized through training methods such as gradient descent. However, there remains a lack of effective and scalable research methods to understand their internal connections and weightings. This leads to the situation where the outputs generated by large models are often difficult for users to fully comprehend. As the "brain" of embodied intelligence, this characteristic further increases the risks of embodied intelligence in practical applications.
At present, in the field of AI security, solutions can only block risks through technical means rather than guide and mitigate them properly, which requires extensive content review and filtering. In terms of data protection, proven security technologies already exist, such as open-source architectures and model watermarking. By watermarking data, it becomes possible to trace the source of risky data and identify the responsible parties, while effectively preventing model jailbreak attacks—attacks that bypass the model's security mechanisms to prompt it to generate or issue prohibited instructions.
In addition, cybersecurity protection remains weak. Hackers can infiltrate large models through Bluetooth, Wi-Fi, and other means, but there are currently no effective technical measures to prevent such attacks. This calls for the development of systematic targeted measures for different application scenarios, the clarification of liability divisions through legal frameworks, and the further improvement of full-chain security guarantees.
Zeng Meiying, head of marketing operations at the National Land Center, admitted to reporters that embodied intelligence will first be implemented in scenarios such as industry, supermarkets, and logistics. However, it will take another 5 to 10 years for it to enter households safely.
|
9
1