Research map / Reinforcement learning / Visual and representation RL
Thin-Shell Object Manipulations With Differentiable Physics Simulations
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.
Introduces ThinShellLab, a fully differentiable simulator for thin-shell and volumetric material manipulation with a benchmark of manipulation and inverse-design tasks, and shows a hybrid CMA-ES plus gradient descent optimizer works best.
Problem
Robots struggle to manipulate diverse thin-shell materials (cloth, paper) because existing simulators lack bending stiffness, bending plasticity, frictional contact with end-effectors, or differentiability, and learning is hampered by sensitive, non-smooth contact-rich dynamics.
Approach
A fully differentiable implicit-Euler simulator with Kirchhoff-Love shell bending, bending plasticity via yield criterion, penalty-based frictional contact, and gradients via implicit function differentiation; combined with a hybrid policy search that initializes with CMA-ES and refines with gradient descent.
Key results
- CMA-ES + GD achieves the best max reward on most manipulation tasks, e.g., Separating 8.24 ± 0.27, Folding-L 4.68 ± 0.17, Pick-Folding 4.74 ± 2.11, versus PPO 6.01 and SAC 7.73 on Separating.
- On inverse design, GD reaches 0.75 on Bouncing and -1.00 on Card, while CMA-ES reaches 4.47 on Sliding versus GD 4.34 there.
- Real-world Separating task was successfully deployed after real-to-sim system identification, whereas a constant-speed heuristic drag failed.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Combine CMA-ES initialization with gradient-based refinement | trajectory optimization · Separating | reward: 8.08 ± 0.18 -> 8.24 ± 0.27 | Helped |
| Combine CMA-ES initialization with gradient-based refinement | trajectory optimization · Following | reward: 3.36 ± 0.70 -> 3.43 ± 0.89 | Helped |
| Combine CMA-ES initialization with gradient-based refinement | trajectory optimization · Folding-U | reward: 3.22 ± 0.04 -> 3.81 ± 0.03 | Helped |
| Combine CMA-ES initialization with gradient-based refinement | trajectory optimization · Folding-L | reward: 3.93 ± 0.80 -> 4.68 ± 0.17 | Helped |
| Combine CMA-ES initialization with gradient-based refinement | trajectory optimization · Pick-Folding | reward: 2.14 ± 2.14 -> 4.74 ± 2.11 | Helped |
| Combine CMA-ES initialization with gradient-based refinement | trajectory optimization · Forming | reward: 3.51 ± 0.26 -> 3.63 ± 0.14 | Helped |
| Combine CMA-ES initialization with gradient-based refinement | trajectory optimization · Lifting | reward: 4.15 -> 2.51 ± 0.39 | Hurt |
| Use CMA-ES instead of GD | inverse design optimization · Sliding | reward: 4.34 ± 0.03 -> 4.47 ± 0.09 | Helped |
| Use GD instead of CMA-ES | inverse design optimization · Bouncing | reward: 0.62 ± 0.04 -> 0.75 ± 0.01 | Helped |
| Use GD instead of CMA-ES | inverse design optimization · Card | reward: -1.02 ± 0.01 -> -1.00 ± 0.04 | Helped |
| Use the hybrid CMA-ES + GD method rather than GD alone | trajectory optimization · Lifting | reward: 4.15 -> 2.51 ± 0.39 | Hurt |
| Use the hybrid CMA-ES + GD method rather than GD alone | trajectory optimization · Separating | reward: 6.80 -> 8.24 ± 0.27 | Helped |
Surprising findings
- Plain gradient descent frequently gets trapped in local optima on folding tasks, and standard RL struggles on Pick-Folding due to sparse rewards.
- Standard RL and CMA-ES alone fail on Lifting, which requires fine-grained coordination of three end-effectors.
Limitations
- Residual sim-to-real gap due to contact model and real-world instability; real-world validation uses only tactile and force sensors.
- Pure gradient-based optimization gets stuck in local optima; RL is sample-inefficient and CMA-ES converges slowly.
Experimental setup
Datasets: real-world tactile and force sensor recordings for system identification
Benchmarks: ThinShellLab manipulation tasks (Lifting, Separating, Following, Folding-U, Folding-L, Pick-Folding, Forming) ThinShellLab inverse design tasks (Sliding, Bouncing, Card)
Metrics: maximum reward within fixed episodes success rate standard deviation across trials
Baselines: PPO SAC CMA-ES gradient descent (Adam)
Builds on: Incremental Potential Contact DiffPD Stable Neo-Hookean Taichi
Closest papers by meaning
- DIFFTACTILE: A Physics-based Differentiable Tactile Simulator for Contact-rich Robotic Manipulation (ICLR 2024 · similarity 0.86)
- GarmentLab: A Unified Simulation and Benchmark for Garment Manipulation (NeurIPS 2024 · similarity 0.85)
- DexGarmentLab: Dexterous Garment Manipulation Environment with Generalizable Policy (NeurIPS 2025 · similarity 0.84)
- Rapidly Adapting Policies to the Real-World via Simulation-Guided Fine-Tuning (ICLR 2025 · similarity 0.83)
- RobotSmith: Generative Robotic Tool Design for Acquisition of Complex Manipulation Skills (NeurIPS 2025 · similarity 0.83)
- Thin-Shell-SfT: Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields (CVPR 2025 · similarity 0.83)
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.