Research map / Point clouds / Point cloud classification

Point Cloud Sequence Encoding for Material-conditioned Graph Network Simulators

NeurIPS 2026 · Point clouds › Point cloud classification · approach: Efficient point sampling architectures · 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 PEACH, an in-context graph network simulator that encodes a few point cloud sequences (via a 4D space-time transformer encoder with auxiliary parameter and SDF losses) into a material latent to predict mesh trajectories for unseen materials without test-time optimization, with zero-shot sim-to-real transfer on a trampoline.

Problem

Graph network simulators assume known material parameters, and mesh trajectories needed for in-context material inference (e.g., MaNGO) are not observable in real-world settings, so adapting learned simulators to new materials from real sensor data is hard.

Approach

Treat each point cloud sequence as a single point cloud in 4D space-time, encode it with a point-patch transformer with Fourier features and attention pooling, aggregate contexts with learned softmax, and condition a trajectory-level MaNGO simulator; add parameter regression and SDF auxiliary losses during training only.

Key results

• PEACH outperforms all point-cloud context baselines by a substantial margin on simulation datasets and beats mesh-based MaNGO on three of four environments.
• On Bending Beam, PEACH surpasses the Oracle baseline in rollout MSE.
• On the real-world Trampoline, PEACH outperforms all baselines including the Oracle, which uses experimentally measured parameters, with zero-shot sim-to-real transfer.

Surprising findings

• PEACH surpasses the Oracle baseline on Bending Beam, which the authors attribute to the latent encoding capturing geometric properties such as thickness alongside material parameters.
• Using only the SDF auxiliary loss remains competitive with using both auxiliary losses.

Limitations

• Requires an initial scene geometry (mesh) to be available.
• Context trajectories must share physical properties with the target process.
• Predicts the full trajectory in a single forward pass, which may limit memory for long sequences and high-resolution meshes.

Experimental setup

Datasets: Trampoline Bending Beam Deforming Block Sheet Deformation

Benchmarks: Trampoline Bending Beam Deforming Block Sheet Deformation Real-world Trampoline

Metrics: MSE point-to-mesh distance

Baselines: MaNGO PSTNet encoder GNN encoder Oracle No Context MeshGraphNets (MGN)

Builds on: MaNGO conditional neural process Point-MAE PSTNet FoundationStereo

Closest papers by meaning

• MaNGO — Adaptable Graph Network Simulators via Meta-Learning (NeurIPS 2025 · similarity 0.89)
• Learning rigid-body simulators over implicit shapes for large-scale scenes and vision (NeurIPS 2024 · similarity 0.85)
• A meshfree exterior calculus for generalizable and data-efficient learning of physics from point clouds (NeurIPS 2026 · similarity 0.85)
• EvoMesh: Adaptive Physical Simulation with Hierarchical Graph Evolutions (ICML 2025 · similarity 0.85)
• Diffusion-Based Hierarchical Graph Neural Networks for Simulating Nonlinear Solid Mechanics (NeurIPS 2025 · similarity 0.85)
• PointInfinity: Resolution-Invariant Point Diffusion Models (CVPR 2024 · similarity 0.84)