Research map / Reinforcement learning / Visual and representation RL

DeformGen: Dynamics-Based Topology Augmentation for Deformable Manipulation Policy Learning

NeurIPS 2026 · Reinforcement learning › Visual and representation RL · approach: Dexterous manipulation transfer · 0 citations (OpenAlex)

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 DeformGen, which augments a single deformable-object demonstration by dynamics-based topological state synthesis in simulation and deformation-field warping of trajectories, improving imitation policies on rope, toy, and cloth tasks.

Problem

Demonstration augmentation for imitation learning works for rigid objects via SE(3) perturbations but breaks for deformable objects, whose valid state variations are high-dimensional and governed by internal dynamics, and whose trajectories cannot be transferred by rigid isometries.

Approach

Generate physically plausible deformed object states by applying localized gripper-driven disturbances and forward-simulating them in PhysTwin, then transfer source demonstrations to each new state with a closed-form KNN inverse-distance deformation field that warps both position and orientation of the end-effector.

Key results

• DeformGen reaches the highest average success rate for ACT (59.00%), SmolVLA (56.50%), and pi0 (56.67%), versus 1 Src. baselines of 1.33-2.50% and SMG* of 27.83-48.17%.
• On rigid-only test states, SMG* (rigid augmentation) is best for ACT, DP, and SmolVLA, while DG is best for pi0 (53.50% average).
• Scaling synthetic data from N=100 to N=750 raises ACT average success from 19.50% to 61.50% and SmolVLA from 36.83% to 63.17%.

Surprising findings

• DeformGen does not beat the rigid-style SMG* baseline on the DP architecture average (37.33 vs 38.00 for DP is close; DG* 41.00 is best) and on some tasks like cloth with DP and SmolVLA, where rigid augmentation is competitive or better.
• Synthesis success varies widely by task (rope 99.5%, toy 60.3%, cloth 39.5%), and the rigid-only test ablation shows topological augmentation can reduce performance on rigid scenarios for ACT and DP.

Limitations

• Evaluated only in simulation (Real2Sim-Eval/PhysTwin) with no real-robot experiments; sim-to-real gap is untested.
• Trajectory synthesis is not universally successful (e.g., 39.5% for cloth, 60.3% for toy), and warped trajectories can fail on large topological changes.
• Only three tasks and a single-arm xArm7 setup; generality to other deformable categories is untested.

Experimental setup

Datasets: single teleoperated demonstration per task (rope routing, toy packing, cloth folding) Real2Sim-Eval simulation assets

Benchmarks: Real2Sim-Eval rope routing Real2Sim-Eval toy packing Real2Sim-Eval cloth folding

Metrics: success rate (%)

Baselines: 1 Src. (single source demo) SoftMimicGen* DeformGen* (local rigid transfer)

Builds on: DemoGen SoftMimicGen PhysTwin Real2Sim-Eval Diffusion Policy

Closest papers by meaning

• DeformMaster: An Interactive Physics-Neural World Model for Deformable Objects from Videos (NeurIPS 2026 · similarity 0.88)
• DynScene: Scalable Generation of Dynamic Robotic Manipulation Scenes for Embodied AI (CVPR 2025 · similarity 0.88)
• Generalizable Domain Adaptation for Sim-and-Real Policy Co-Training (NeurIPS 2025 · similarity 0.87)
• Geometry-aware RL for Manipulation of Varying Shapes and Deformable Objects (ICLR 2025 · similarity 0.87)
• DynaRend: Learning 3D Dynamics via Masked Future Rendering for Robotic Manipulation (NeurIPS 2025 · similarity 0.87)
• Predictive Inverse Dynamics Models are Scalable Learners for Robotic Manipulation (ICLR 2025 · similarity 0.87)