Research map / Reinforcement learning / Visual and representation RL
BodyGen: Advancing Towards Efficient Embodiment Co-Design
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes BodyGen, an RL-based embodiment co-design framework using topology-aware self-attention (MoSAT) with topology position encoding (TopoPE) and a decoupled temporal credit assignment (Enhanced-TCA) to jointly optimize morphology and control efficiently.
Problem
Embodiment co-design, jointly optimizing robot morphology and control policy, is inefficient due to combinatorial morphology search spaces and intricate morphology-control dependencies, plus unbalanced reward signals between design and control stages.
Approach
Use an autoregressive, joint-level self-attention transformer (MoSAT) with a hash-based path-to-root topology position encoding so morphology is represented compactly and shared across similar bodies, and decouple design and control advantage estimation (modified GAE) so both stages receive balanced credit under PPO.
Key results
- BodyGen achieves an average 60.03% performance improvement over the strongest baseline across ten co-design tasks.
- BodyGen per-agent model has 1.43M parameters versus 19.64M for Transform2Act.
- Removing MoSAT drops Crawler return from 10381.96 to 818.92 for BodyGen.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace TopoPE with Traversal PE (ablation) | BodyGen · 10 co-design environments | episode reward: 10381.96 ± 353.97 vs 8582.24 ± 987.44 (Crawler); TopoPE is higher in all 10 environments | Helped |
| remove TopoPE (ablation) | BodyGen · 9 of 10 co-design environments | episode reward: 10381.96 ± 353.97 vs 7490.83 ± 267.70 (Crawler); 1305.17 ± 15.25 vs 1371.20 ± 30.74 (Swimmer) | Hurt |
| remove MoSAT centralized attention (ablation) | BodyGen · 10 co-design environments | episode reward: 10381.96 ± 353.97 vs 818.92 ± 57.78 (Crawler) | Hurt |
| remove Enhanced-TCA and use original PPO (ablation) | BodyGen · 10 co-design environments | episode reward: 10381.96 ± 353.97 vs 4994.44 ± 160.14 (Crawler) | Hurt |
| use BodyGen instead of the strongest listed baseline | embodiment co-design · 10 co-design environments | episode reward: +60.03% average performance improvement | Helped |
Surprising findings
- Without TopoPE, Glider-Hard return collapses to 109.48 while Swimmer slightly improves without TopoPE (1371.20 vs 1305.17).
Limitations
- Only evaluated in simulation (MuJoCo); no transfer to real physical robots.
- Reward-driven RL focuses on control effectiveness and does not model rich perception/execution of biological systems.
Experimental setup
Datasets: MuJoCo co-design environments (Crawler, TerrainCrosser, Cheetah, Swimmer, Glider, Walker)
Benchmarks: Crawler TerrainCrosser Cheetah Swimmer Glider-Regular Glider-Medium Glider-Hard Walker-Regular Walker-Medium Walker-Hard
Metrics: episodic return (mean ± standard error over four seeds) convergence speed model parameter count
Baselines: NGE Transform2Act UMC-Message* UMC-Message BodyGen w/o MoSAT BodyGen w/o Enhanced-TCA
Builds on: PPO GAE Transform2Act UMC-Message NGE
Closest papers by meaning
- Embracing Evolution: A Call for Body-Control Co-Design in Embodied Humanoid Robot (NeurIPS 2026 · similarity 0.90)
- Convergent Functions, Divergent Forms (NeurIPS 2025 · similarity 0.88)
- DittoGym: Learning to Control Soft Shape-Shifting Robots (ICLR 2024 · similarity 0.86)
- MorphVAE: Advancing Morphological Design of Voxel-Based Soft Robots with Variational Autoencoders (AAAI 2024 · similarity 0.86)
- Debate2Create: Robot Co-design via Multi-Agent LLM Debate (ICML 2026 · similarity 0.85)
- Hierarchical World Models as Visual Whole-Body Humanoid Controllers (ICLR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.