BodyGen: A Bio-Inspired Framework for Rapid Co-Evolution of Robot Morphology and Control

In a groundbreaking study accepted as a Spotlight paper at ICLR 2025, researchers from Ant Group and Tsinghua University present BodyGen, a novel framework enabling robots to autonomously evolve both their physical forms and control strategies. Inspired by biological evolution, BodyGen combines reinforcement learning (RL) and neural networks to optimize robot morphology and behavior in simulated environments, achieving a 60% performance improvement over state-of-the-art baselines with just 1.43 million parameters.

Why Autonomous Robot Evolution Matters

Traditional robot design relies on human experts to manually iterate on physical structures and control algorithms—a time-consuming process ill-suited for dynamic environments. BodyGen addresses this by mimicking natural selection: robots evolve through trial-and-error interactions, refining both their "bodies" (morphology) and "brains" (control policies) simultaneously.

The BodyGen Framework: Key Innovations

BodyGen introduces three core technologies to tackle morphological and control co-design:

1. TopoPE: Topology-Aware Position Encoding

Like a biological nervous system, TopoPE assigns unique "intelligent labels" to robot limbs. By encoding the topological relationship between body parts (e.g., "this joint is connected to the left leg"), TopoPE ensures the AI understands structural changes during evolution, enabling knowledge transfer across different robot shapes.

2. MoSAT: Modular Self-Attention Transformer

Acting as a centralized "brain," MoSAT processes sensory data from all robot joints through a Transformer architecture:

3. Temporal Credit Assignment with Hierarchical GAE

BodyGen uses a modified Proximal Policy Optimization (PPO) algorithm to balance rewards for morphological decisions (e.g., adding a limb) and control actions (e.g., moving joints). A hierarchical Generalized Advantage Estimation (GAE) ensures delayed rewards from morphological changes are properly credited, preventing short-sighted design choices.

Performance and Validation

In simulations across 10 diverse environments (e.g., terrain traversal, swimming), BodyGen outperformed baselines like Transform2Act and NGE by 60.03% in adaptive efficiency. Key findings include:

Applications and Future Directions

BodyGen opens new possibilities for:

The team plans to extend BodyGen to real-world robotics through physics-to-reality transfer learning, aiming to create self-optimizing robots capable of lifelong adaptation.