14/2026-03
The Rise of Embodied Intelligence: How AI Enables Robots to Achieve 'Autonomous Evolution'?
At the 2025 Spring Festival Gala, the scene of Unitree Robotics' robots spinning handkerchiefs and dancing Yangko with humans was impressive.
'AI Godfather' Nvidia's Jensen Huang has repeatedly emphasized that the next wave of AI is embodied intelligence. This year, 'embodied intelligence' was written into the government work report for the first time.
Embodied Intelligence is a concept involving artificial intelligence, cognitive science, and robotics, referring to the integration of AI into physical entities like robots, endowing them with the ability to perceive, learn, and dynamically interact with the environment.
Simply put, it's about enabling robots to accurately perceive and flexibly respond in the real world, much like humans do.
How can this be achieved? Perhaps giving robots autonomous evolutionary capabilities similar to humans is a viable direction.
Today, we present a guide to a Spotlight technical paper from ICLR2025. The first author of this paper, Lu Haofei, and the second author, Wu Zhe, are master's and doctoral students respectively at the Human-Computer Interaction Laboratory of Tsinghua University, with collaborators including Li Jianshu, a senior algorithm expert at Ant Digital Technologies.
The research team of this paper has provided an affirmative answer through experiments: by combining reinforcement learning and deep neural network technologies, robots can automatically evolve optimal morphologies and control strategies adapted to the current environment in a very short time, offering a new approach for accelerating the evolution of embodied intelligence.
The research team comes from Tsinghua University and Ant Digital Technologies, and they named this algorithm framework BodyGen.
Actual test results show that BodyGen can achieve a 60% performance improvement.
Currently, the code for this framework has been open-sourced to the GitHub repository >> Click here to visit, and everyone is welcome to experience it~
01 Why Robots Need 'Autonomous Evolution'
Natural organisms have acquired perfect body structures and environmental interaction capabilities adapted to their environments through millions of years of evolution. However, robot design not only requires extensive human expert knowledge but also demands numerous experiments, designs, and iterations for specific environments.
Inspired by natural biology, scientists have proposed morphology-control co-design technology: allowing robots to mimic the biological evolution process, evolving their own morphology (physical structure, joint parameters) while optimizing control strategies (the brain) to adapt to complex environments.
However, this field has long faced two major challenges: first, the morphology search space is enormous, making it difficult to exhaust all possible robot morphologies within limited time; second, robot morphology and control strategies are deeply coupled, requiring substantial computational resources to evaluate each candidate morphology design.
The research team from Tsinghua University and Ant Digital Technologies proposed the BodyGen framework, utilizing reinforcement learning to achieve efficient end-to-end morphology-control co-design.
In this work, the morphology design process is divided into two consecutive stages: the morphology design stage and the environmental interaction stage.
In the morphology design stage: the research introduces a Transformer (GPT-Style) to autoregressively construct the robot's physical structure and optimize parameters;
In the environmental interaction stage: a Transformer (Bert-Style) is also used to centrally process each joint's information of the robot and send it to the corresponding joint motors, interacting with the environment and receiving feedback.
After several rounds of iteration, BodyGen can quickly generate the optimal robot morphology and corresponding control strategies for the current environment.
02 Interpretation of BodyGen's Three Key Technical Points
1. The robot's 'body perception' system: Topological Position Encoder (TopoPE);
TopoPE acts like the robot's 'body perception' system by attaching 'smart labels' to each part of the robot. Regardless of how the robot's shape changes, these labels help the AI understand 'this is a leg,' 'this is an arm,' etc. This way, even when the robot's morphology changes, the AI can quickly adapt and control the new body structure.
2. The robot's 'brain central': Transformer-based centralized neural hub processing network MoSAT;
MoSAT acts like the robot's 'brain central,' and its working method is very similar to the human brain.
Information collection: First, collect information from various parts of the robot, such as position, speed, etc.
Central processing: All information is communicated and processed in the 'brain' (Transformer network).
Instruction sending: The processed information is transformed into specific action commands, telling the robot how to move.
3. Reward allocation mechanism: Co-design under temporal credit assignment mechanism.
BodyGen allows AI to be responsible for two things simultaneously: designing the robot's body and controlling the robot's actions.
Design actions: AI can 'grow' new limbs for the robot, 'cut off' unnecessary parts, or maintain the existing structure.
Control actions: AI learns how to control each joint of the robot to complete tasks (such as walking, jumping).
Designing a good robot morphology may require a long time to know whether the effect is good. For example, after designing long legs, one has to wait until the robot learns to walk to know whether this design is reasonable.
Through a special 'reward allocation mechanism,' BodyGen enables AI to reasonably evaluate its design decisions and not abandon potentially excellent designs due to unobvious short-term effects.
03 Test Results: 60% Performance Improvement
In 10 different task environments (such as crawling, terrain crossing, swimming, etc.), the robot morphologies generated by BodyGen achieved a 60.03% increase in environmental adaptability scores compared to existing optimal methods (such as Transform2Act, NGE, etc.).
BodyGen has an average parameter count of 1.43M, making it more lightweight compared to other baseline algorithms. BodyGen's compact design significantly reduces computational costs and storage requirements while maintaining generation capabilities, improving training stability and efficiency.
This efficient model architecture not only reduces resource consumption during training but also improves response speed during the inference stage, making it more suitable for resource-constrained environments, and can even perform efficient inference on a laptop using CPU.
In the future, the team plans to promote the application of BodyGen in real-world scenarios through physical simulation transfer technology. With the improvement of computing power, this framework is expected to become an important path toward achieving general embodied intelligence, enabling robots to continuously optimize morphology and behavioral strategies through perception-action loops, gradually achieving self-design and automatic evolution.
'AI Godfather' Nvidia's Jensen Huang has repeatedly emphasized that the next wave of AI is embodied intelligence. This year, 'embodied intelligence' was written into the government work report for the first time.
Embodied Intelligence is a concept involving artificial intelligence, cognitive science, and robotics, referring to the integration of AI into physical entities like robots, endowing them with the ability to perceive, learn, and dynamically interact with the environment.
Simply put, it's about enabling robots to accurately perceive and flexibly respond in the real world, much like humans do.
How can this be achieved? Perhaps giving robots autonomous evolutionary capabilities similar to humans is a viable direction.
Today, we present a guide to a Spotlight technical paper from ICLR2025. The first author of this paper, Lu Haofei, and the second author, Wu Zhe, are master's and doctoral students respectively at the Human-Computer Interaction Laboratory of Tsinghua University, with collaborators including Li Jianshu, a senior algorithm expert at Ant Digital Technologies.
The research team of this paper has provided an affirmative answer through experiments: by combining reinforcement learning and deep neural network technologies, robots can automatically evolve optimal morphologies and control strategies adapted to the current environment in a very short time, offering a new approach for accelerating the evolution of embodied intelligence.
The research team comes from Tsinghua University and Ant Digital Technologies, and they named this algorithm framework BodyGen.
Actual test results show that BodyGen can achieve a 60% performance improvement.
Currently, the code for this framework has been open-sourced to the GitHub repository >> Click here to visit, and everyone is welcome to experience it~
01 Why Robots Need 'Autonomous Evolution'
Natural organisms have acquired perfect body structures and environmental interaction capabilities adapted to their environments through millions of years of evolution. However, robot design not only requires extensive human expert knowledge but also demands numerous experiments, designs, and iterations for specific environments.
Inspired by natural biology, scientists have proposed morphology-control co-design technology: allowing robots to mimic the biological evolution process, evolving their own morphology (physical structure, joint parameters) while optimizing control strategies (the brain) to adapt to complex environments.
However, this field has long faced two major challenges: first, the morphology search space is enormous, making it difficult to exhaust all possible robot morphologies within limited time; second, robot morphology and control strategies are deeply coupled, requiring substantial computational resources to evaluate each candidate morphology design.
The research team from Tsinghua University and Ant Digital Technologies proposed the BodyGen framework, utilizing reinforcement learning to achieve efficient end-to-end morphology-control co-design.
In this work, the morphology design process is divided into two consecutive stages: the morphology design stage and the environmental interaction stage.
In the morphology design stage: the research introduces a Transformer (GPT-Style) to autoregressively construct the robot's physical structure and optimize parameters;
In the environmental interaction stage: a Transformer (Bert-Style) is also used to centrally process each joint's information of the robot and send it to the corresponding joint motors, interacting with the environment and receiving feedback.
After several rounds of iteration, BodyGen can quickly generate the optimal robot morphology and corresponding control strategies for the current environment.
02 Interpretation of BodyGen's Three Key Technical Points
1. The robot's 'body perception' system: Topological Position Encoder (TopoPE);
TopoPE acts like the robot's 'body perception' system by attaching 'smart labels' to each part of the robot. Regardless of how the robot's shape changes, these labels help the AI understand 'this is a leg,' 'this is an arm,' etc. This way, even when the robot's morphology changes, the AI can quickly adapt and control the new body structure.
2. The robot's 'brain central': Transformer-based centralized neural hub processing network MoSAT;
MoSAT acts like the robot's 'brain central,' and its working method is very similar to the human brain.
Information collection: First, collect information from various parts of the robot, such as position, speed, etc.
Central processing: All information is communicated and processed in the 'brain' (Transformer network).
Instruction sending: The processed information is transformed into specific action commands, telling the robot how to move.
3. Reward allocation mechanism: Co-design under temporal credit assignment mechanism.
BodyGen allows AI to be responsible for two things simultaneously: designing the robot's body and controlling the robot's actions.
Design actions: AI can 'grow' new limbs for the robot, 'cut off' unnecessary parts, or maintain the existing structure.
Control actions: AI learns how to control each joint of the robot to complete tasks (such as walking, jumping).
Designing a good robot morphology may require a long time to know whether the effect is good. For example, after designing long legs, one has to wait until the robot learns to walk to know whether this design is reasonable.
Through a special 'reward allocation mechanism,' BodyGen enables AI to reasonably evaluate its design decisions and not abandon potentially excellent designs due to unobvious short-term effects.
03 Test Results: 60% Performance Improvement
In 10 different task environments (such as crawling, terrain crossing, swimming, etc.), the robot morphologies generated by BodyGen achieved a 60.03% increase in environmental adaptability scores compared to existing optimal methods (such as Transform2Act, NGE, etc.).
BodyGen has an average parameter count of 1.43M, making it more lightweight compared to other baseline algorithms. BodyGen's compact design significantly reduces computational costs and storage requirements while maintaining generation capabilities, improving training stability and efficiency.
This efficient model architecture not only reduces resource consumption during training but also improves response speed during the inference stage, making it more suitable for resource-constrained environments, and can even perform efficient inference on a laptop using CPU.
In the future, the team plans to promote the application of BodyGen in real-world scenarios through physical simulation transfer technology. With the improvement of computing power, this framework is expected to become an important path toward achieving general embodied intelligence, enabling robots to continuously optimize morphology and behavioral strategies through perception-action loops, gradually achieving self-design and automatic evolution.