Amortized NeuralSDF-Mesh Collision Detection for Robotic Contact Simulation
Jinhee Yun, Jeongmin Lee, Sunkyung Park, Dongjun Lee
AI summary
Problem
Neural Signed Distance Functions (NeuralSDFs) offer high-fidelity, memory-efficient geometry representation but are too slow for real-time robotics because each distance query requires a neural network forward pass, and iterative collision solvers multiply this latency.
Approach
The authors formulate NeuralSDF-mesh collisions as a constrained optimization problem, derive Karush-Kuhn-Tucker optimality conditions, and train an auto-decoder network to directly predict contact points in a single pass, bypassing iterative numerical solvers.
Key results
- Reduces per-triangle collision detection time to ~0.038 ms, a 1000x speedup over iterative baselines
- Maintains high contact optimality and accuracy comparable to iterative optimization methods
- Lowers memory footprint by replacing high-resolution voxel grids with continuous neural representations
- Enables category-level generalization to unseen objects without retraining
Why it matters
It removes the computational bottleneck of NeuralSDF queries, enabling real-time, high-fidelity contact simulation for complex robotic manipulation and assembly tasks.
Abstract
Collision detection is a fundamental problem in robotics, but handling collisions between non-convex objects remains challenging. A common approach for representing non- convex geometry is a signed distance function (SDF). Voxel- based SDF (VoxelSDF) enables fast distance queries but suffers from discretization artifacts and high memory costs. Neural implicit SDF (NeuralSDF) provides a continuous and memory- efficient representation with generalization, yet their slow query speed has limited their use in collision detection. To overcome these limitations, this paper proposes a novel amortized Neu- ralSDF–mesh collision detection framework. NeuralSDF–mesh collisions are formulated as a constrained optimization problem at the triangle level, and the Karush–Kuhn–Tucker conditions are derived to enable the amortization. A learning-based amortized optimization directly predicts collisions in a single forward pass, eliminating iterative optimization procedures. The amortized model adopts an auto-decoder architecture, extending the advantages of NeuralSDF in memory efficiency and category-level generalization to collision detection. Experi- ments demonstrate substantial speedups over baseline methods while maintaining comparable contact quality and reduced memory usage. The proposed approach also exhibits category- level generalization to unseen objects and can be applied to various robotic simulation scenarios.