Research map / Reinforcement learning / Visual and representation RL
Generating Freeform Endoskeletal Robots
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.
Generates freeform endoskeletal soft robots by encoding voxel body plans (soft elastic voxels plus rigid bone voxels with joints) in a VAE latent space, co-optimizing morphology with a universal PPO controller and CMA-ES evolution in a new simulator.
Problem
Automatic design of robots is limited to fully rigid jointed or fully soft jointless bodies; combining soft tissue and internal rigid skeletons with joints is hard to simulate and optimize in a structured, open-ended way.
Approach
Model endoskeletal body plans as voxel grids of elastic and rigid cells, learn a VAE latent embedding trained on unlimited procedurally generated multi-star synthetic designs, and evolve latent codes with CMA-ES while training a single graph-transformer universal controller via PPO across the population.
Key results
- Evolved designs significantly outperform initial random designs in all four task environments, with frozen evolved populations beating frozen initial populations under both optimized-RL and from-scratch retraining.
- Random morphological search performed significantly worse than CMA-ES evolution in the Upright Locomotion task.
- Training 64 individual controllers for 30 epochs took 44.8 hours on 4 H100 GPUs, whereas a universal controller could train for 451 epochs in the same time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use evolutionary strategies instead of random morphological search | Endoskeletal robot design optimization · Upright Locomotion | fitness | Helped |
| Evolve the designs rather than use the initial randomly generated designs | Universal controller training · Flat Ground, Upright Locomotion, Potholes, and Mountain Range | fitness | Helped |
| Replace discrete actions with continuous actions (ablation) | Universal controller · Upright Locomotion | fitness | No clear effect |
| Replace the universal controller with separate individual controllers (ablation) | Controller training across 64 designs · Upright Locomotion | fitness | No clear effect |
Surprising findings
- Continuous action spaces were viable, but a discrete action set greatly simplified and stabilized policy training.
- On flat ground, evolution converged to two-jointed snakes and failed to escape that local optimum for legs even after 100 generations.
Limitations
- Designs were not physically fabricated; results are simulation-only with no sim2real transfer tested.
- Rigid body collisions were not modeled, so external rigid structures such as horns or claws are impossible, and fluids are not modeled.
- Skeletons are assumed internal, and the simulator is restricted to terrestrial environments.
Experimental setup
Datasets: procedurally generated synthetic multi-star endoskeletal body plans
Benchmarks: Flat Ground Upright Locomotion Potholes Mountain Range
Metrics: peak fitness (displacement reward) net displacement in meters reconstruction accuracy
Baselines: random morphological search individual per-design controllers continuous action space
Builds on: Proximal Policy Optimization CMA-ES Compositional Pattern Producing Networks Voxception-ResNet Euler-Bernoulli beam voxel simulation
Closest papers by meaning
- MorphVAE: Advancing Morphological Design of Voxel-Based Soft Robots with Variational Autoencoders (AAAI 2024 · similarity 0.85)
- Embracing Evolution: A Call for Body-Control Co-Design in Embodied Humanoid Robot (NeurIPS 2026 · similarity 0.85)
- Convergent Functions, Divergent Forms (NeurIPS 2025 · similarity 0.84)
- BodyGen: Advancing Towards Efficient Embodiment Co-Design (ICLR 2025 · similarity 0.84)
- EnliveningGS: Active Locomotion of 3DGS (CVPR 2025 · similarity 0.83)
- DiffeoMorph: Learning to Morph 3D Shapes Using Differentiable Agent-Based Simulations (NeurIPS 2026 · similarity 0.82)
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.