Research map / Probabilistic and scientific modeling / Variational and generative sampling
MorphVAE: Advancing Morphological Design of Voxel-Based Soft Robots with Variational Autoencoders
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 MorphVAE, a multi-task VAE trained bootstrapping-style via "continuous natural selection" sampling to design voxel-based soft robots across locomotion and manipulation tasks, with sample-efficiency and diversity gains over EA baselines.
Problem
Free-form voxel-based soft robot morphology design has a vast combinatorial search space; existing evolutionary and probabilistic generative approaches are sample-inefficient and focus on rigid robots and locomotion.
Approach
A task-conditioned VAE trained on a morphology pool with fitness-proportional (temperature-controlled) resampling, so design experience is shared across tasks and generations, plus an exploration-exploitation rebalancing step to boost diversity.
Key results
- MorphVAE-H significantly outperforms competitors in four of six Evolution Gym tasks in maximal fitness, and is competitive on Walker-v0 and Catcher-v0.
- Average diversity rank across six tasks: MorphVAE-B obtained 2.17 for k=10 (best among methods) and 1.83 for k=5 for MorphVAE-H... MorphVAE-B ranked 2.17 (k=10) and 2.17 (k=5) per the reported ordering.
- MorphVAE-H's fitness distributions lie higher than baselines across all tasks, per boxplots.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MorphVAE-H instead of the competitive baselines | Soft-robot design in Evolution Gym · Carrier-v0, Pusher-v0, Climber-v0, UpStepper-v0, Walker-v0, and Catcher-v0 | maximal fitness | Helped |
| Use MorphVAE-H instead of the competitive baselines | Soft-robot design in Evolution Gym · Six Evolution Gym tasks | fitness distribution | Helped |
| Use MorphVAE-B instead of the competitive baselines | Soft-robot design in Evolution Gym · Six Evolution Gym tasks | robot diversity | Helped |
| Use MorphVAE-B instead of MorphVAE-H, with lower selection temperature and exploration-exploitation rebalancing (ablation) | MorphVAE soft-robot design · Six Evolution Gym tasks | robot diversity: 2.17 average rank at k=10; 1.83 average rank at k=5, compared with 5.83 and 5.17 for MorphVAE-H | Helped |
| Use MorphVAE-B instead of MorphVAE-H, with lower selection temperature and exploration-exploitation rebalancing (ablation) | MorphVAE soft-robot design · Six Evolution Gym tasks | maximal fitness | Mixed |
| Set selection temperature τ to 1.5 (MorphVAE-H) rather than 0.7 (MorphVAE-B) (ablation) | MorphVAE · Six Evolution Gym tasks | optimization efficiency | Helped |
| Add exploration-exploitation rebalancing and use selection temperature τ=0.7 (MorphVAE-B) (ablation) | MorphVAE · Six Evolution Gym tasks | robot diversity | Helped |
Surprising findings
- Random-like GA performs better than MorphVAE on Walker-v0 and Catcher-v0, where MorphVAE plateaus after reaching competitive convergence speed.
Limitations
- Slightly poorer performance on extremely simple (Walker-v0) and extremely hard (Catcher-v0) tasks.
- Uses fixed 5x5 bounding box to keep search space tractable.
- Relies on a fixed PPO controller, which the authors suggest may need adaptive complexity.
Experimental setup
Datasets: Evolution Gym tasks (Walker-v0, Climber-v0, UpStepper-v0, Catcher-v0, Carrier-v0, Pusher-v0)
Benchmarks: Evolution Gym Walker-v0 Evolution Gym Climber-v0 Evolution Gym UpStepper-v0 Evolution Gym Catcher-v0 Evolution Gym Carrier-v0 Evolution Gym Pusher-v0
Metrics: maximal fitness fitness distribution diversity
Baselines: CPPN-NEAT Bayesian Optimization Genetic Algorithm Speciated Evolver roboGAN
Builds on: Variational Autoencoder Proximal Policy Optimization Speciated Evolver roboGAN Evolution Gym
Closest papers by meaning
- MorphoGen: Free-Form XML Evolution for Robot Morphology Design (NeurIPS 2026 · similarity 0.91)
- SoftZoo: A Soft Robot Co-design Benchmark For Locomotion In Diverse Environments (ICLR 2023 · similarity 0.89)
- DittoGym: Learning to Control Soft Shape-Shifting Robots (ICLR 2024 · similarity 0.87)
- Convergent Functions, Divergent Forms (NeurIPS 2025 · similarity 0.87)
- BodyGen: Advancing Towards Efficient Embodiment Co-Design (ICLR 2025 · similarity 0.86)
- Generating Freeform Endoskeletal Robots (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.