Tsinghua and Ant Group Unveil BodyGen: Robot Performance Up 60%

In a significant leap for embodied AI, researchers from Tsinghua University and Ant Group have introduced BodyGen, an innovative algorithm framework that boosts robot performance by an impressive 60%. This breakthrough, detailed in their ICLR 2025 paper, combines reinforcement learning and deep neural networks to enable robots to rapidly adapt their morphology and control strategies to their environment. The team has also open-sourced the code on GitHub, making it accessible to the broader scientific community.

Addressing Traditional Challenges

Traditional robot design has long faced hurdles such as heavy reliance on expert knowledge, iterative experimentation for specific environments, and the complex interplay between morphology and control strategies. The BodyGen framework tackles these issues through a two-stage process: morphology design and environmental interaction. In the first stage, a Transformer-based autoregressive model (GPT-Style) constructs and optimizes robot body structure parameters. In the second stage, another Transformer (Bert-Style) processes joint information to achieve interactive feedback with the environment.

Core Technologies of BodyGen

The framework incorporates three key technologies:

1. TopoPE (Topological Positional Encoder): Acts as the robot's "body awareness" system, enabling rapid adaptation to morphological changes.
2. MoSAT (Transformer-based Central Brain): Handles information processing and command issuance.
3. Specialized Reward Allocation Mechanism: Allows the AI to rationally evaluate design decisions.

Impressive Test Results

In tests across 10 diverse tasks—including crawling and swimming—robots generated by BodyGen achieved a 60.03% higher adaptability score compared to existing state-of-the-art methods. Additionally, with a parameter count of just 1.43M, BodyGen is significantly more lightweight, making it ideal for resource-constrained environments.

The research team plans to promote its application in real-world scenarios, positioning BodyGen as a key enabler for the development of general-purpose embodied AI. For more details or to explore the code, visit GitHub.

Key Points

1. BodyGen enhances robot performance by 60% through optimized morphology and control strategies.
2. The framework addresses traditional challenges in robotics with a two-stage process: morphology design and environmental interaction.
3. Core technologies include TopoPE, MoSAT, and a specialized reward allocation mechanism.
4. Test results show significant improvements in adaptability across diverse tasks.
5. The lightweight design (1.43M parameters) makes it suitable for resource-constrained environments.