BodyGen: A Bio‑Inspired Embodied Co‑Design Framework for Autonomous Robot Evolution

At the ICLR 2025 conference, a Spotlight paper titled "BodyGen" was accepted, authored by a joint team from Ant Financial’s Tianji Lab and Tsinghua University. The work introduces a novel embodied co‑design framework that simultaneously evolves robot morphology and control strategies.

The authors investigate whether robots can autonomously evolve like biological organisms. By integrating reinforcement learning with deep neural networks, BodyGen rapidly discovers optimal robot shapes and corresponding control policies tailored to the current environment.

The implementation of BodyGen has been open‑sourced on GitHub ( https://github.com/GenesisOrigin/BodyGen ), allowing the community to reproduce and extend the experiments.

ICLR 2025 received 11,672 submissions, with roughly 5.1 % selected for Spotlight or oral presentation, highlighting the significance of this work.

Robotic design traditionally suffers from two major challenges: an enormous search space for possible morphologies and a tight coupling between body structure and control policy, which makes exhaustive evaluation computationally expensive.

Inspired by biological evolution, the authors propose a co‑design approach that jointly optimizes the robot’s “body” (shape, joint parameters) and its “brain” (control policy) using a transformer‑based pipeline.

BodyGen’s three core technical contributions are: (1) TopoPE, a lightweight topology‑aware positional encoder that tags each robot part with a learned embedding, enabling the model to recognize and adapt to structural changes; (2) MoSAT, a centralized neural hub that processes all joint information via transformer networks (GPT‑style for shape generation and BERT‑style for control), performing encoding, centralized processing, and decoding; and (3) a temporal credit‑assignment mechanism that balances reward distribution between morphology design and control learning.

TopoPE maps the path from each limb to the root body into a unique embedding using hash‑based encoding, which mitigates index‑shift issues during shape evolution and facilitates knowledge sharing across similar robot designs.

MoSAT treats the robot’s joint data as a sequence, applying a transformer to enable point‑to‑point communication among joints, followed by decoding to produce motor torque commands.

Training employs Proximal Policy Optimization (PPO) with an enhanced Generalized Advantage Estimation (GAE) that hierarchically allocates rewards, allowing the agent to receive balanced feedback for both shape and control actions.

Experiments on three base topologies (linear, bipedal, quadrupedal) across ten diverse simulated environments (e.g., crawling, terrain traversal, swimming) show that BodyGen achieves an average 60.03 % performance improvement over state‑of‑the‑art baselines such as Transform2Act and NGE, while using only 1.43 M parameters and running efficiently on a single CPU.

Potential applications include rapid environment‑adaptive robot design, biomimetic robot research (designing limbs, fins, wings), and virtual character motion generation for games and animation.

Future work aims to transfer BodyGen to real‑world scenarios via physical simulation migration, advancing general embodied intelligence where robots continuously refine their morphology and behavior through perception‑action loops.