Beijing, China – In a significant stride towards advancing embodied intelligence, a joint team from Ant Group Digital Technologies and Tsinghua University has developed BodyGen, a novel framework enabling robots to autonomously evolve their morphology and control strategies. The research, accepted as a Spotlight paper at the prestigious International Conference on Learning Representations (ICLR) 2025, explores the intriguing question of whether robots can, like biological organisms, adapt and evolve to their environments.
The team’s affirmative answer, achieved through the integration of reinforcement learning and deep neural network technologies, has garnered significant attention within the AI community. BodyGen demonstrates the ability to rapidly evolve optimal robot forms and control policies tailored to specific environments, offering a fresh perspective on accelerating the evolution of embodied intelligence.
Nature has spent millions of years evolving the perfect body structures and environmental interaction capabilities, explains Lu Haofei, the first author of the paper and a Master’s student at Tsinghua University’s Department of Computer Science. Wu Zhe, the second author and a Ph.D. candidate in the same department, adds, Our research aims to replicate this evolutionary process in robots, allowing them to adapt and optimize their physical form and behavior for specific tasks.
Professor Xing Junliang, the corresponding author and a long-time researcher in perception and game-theoretic decision-making, emphasizes the broader implications of the work. This research opens up new avenues for developing more adaptable and robust robots that can operate effectively in complex and dynamic environments, he stated. Professor Xing’s research focuses on multi-agent systems, reinforcement learning, and intelligent decision-making.
The BodyGen framework allows for the co-design of both the robot’s body and its control system, enabling a synergistic optimization process. This is a crucial step towards creating truly autonomous robots that can learn and adapt in real-time.
The paper, titled BodyGen: Advancing Towards Efficient Embodiment Co-Design, is available on arXiv at https://arxiv.org/abs/2503.00533. The team has also open-sourced the code on GitHub at https://github.com/GenesisOrigin/BodyGen, encouraging further research and development in this exciting field.
The ICLR 2025 conference received a total of 11,672 submissions, making the selection of BodyGen as a Spotlight paper a testament to the significance and potential impact of this research. The framework promises to be a valuable tool for researchers and engineers working on the next generation of intelligent robots, paving the way for more adaptable, efficient, and autonomous machines.
References:
- Lu, H., Wu, Z., & Xing, J. (2025). BodyGen: Advancing Towards Efficient Embodiment Co-Design. arXiv preprint arXiv:2503.00533. Retrieved from https://arxiv.org/abs/2503.00533
- GitHub Repository: https://github.com/GenesisOrigin/BodyGen
Views: 0